Automatic Screening and Classification of Diabetic ...

Automatic Screening and Classification of Diabetic Retinopathy Eye Fundus Images Rahim, S.S. Submitted version deposited in Coventry University’s Institutional Repository Original citation: Rahim, S.S. (2016) Automatic Screening and Classification of Diabetic Retinopathy Eye Fundus Images. Unpublished PhD Thesis. Coventry: Coventry University Copyright © and Moral Rights are retained by the author. A copy can be downloaded for personal non-commercial research or study, without prior permission or charge. This item cannot be reproduced or quoted extensively from without first obtaining permission in writing from the copyright holder(s). The content must not be changed in any way or sold commercially in any format or medium without the formal permission of the copyright holders. Some materials have been removed from this thesis due to third party copyright. Pages where material has been removed are clearly marked in the electronic version. The unabridged version of the thesis can be viewed at the Lanchester Library, Coventry University.

Automatic Screening and

Classification of Diabetic

Retinopathy Eye Fundus Images

Sarni Suhaila Rahim

June 2016

By

A thesis submitted in partial fulfilment of the University’s

requirements for the Degree of Doctor of Philosophy

i

ABSTRACT

Diabetic Retinopathy (DR) is a disorder of the retinal vasculature. It develops to

some degree in nearly all patients with long-standing diabetes mellitus and can result in

blindness. Screening of DR is essential for both early detection and early treatment.

This thesis aims to investigate automatic methods for diabetic retinopathy detection and

subsequently develop an effective system for the detection and screening of diabetic

retinopathy.

The presented diabetic retinopathy research involves three development stages.

Firstly, the thesis presents the development of a preliminary classification and screening

system for diabetic retinopathy using eye fundus images. The research will then focus

on the detection of the earliest signs of diabetic retinopathy, which are the

microaneurysms. The detection of microaneurysms at an early stage is vital and is the

first step in preventing diabetic retinopathy. Finally, the thesis will present decision

support systems for the detection of diabetic retinopathy and maculopathy in eye fundus

images. The detection of maculopathy, which are yellow lesions near the macula, is

essential as it will eventually cause the loss of vision if the affected macula is not treated

in time.

An accurate retinal screening, therefore, is required to assist the retinal screeners

to classify the retinal images effectively. Highly efficient and accurate image processing

techniques must thus be used in order to produce an effective screening of diabetic

retinopathy. In addition to the proposed diabetic retinopathy detection systems, this

thesis will present a new dataset, and will highlight the dataset collection, the expert

diagnosis process and the advantages of the new dataset, compared to other public eye

fundus images datasets available. The new dataset will be useful to researchers and

practitioners working in the retinal imaging area and would widely encourage

comparative studies in the field of diabetic retinopathy research. It is envisaged that the

proposed decision support system for clinical screening would greatly contribute to and

assist the management and the detection of diabetic retinopathy. It is also hoped that the

developed automatic detection techniques will assist clinicians to diagnose diabetic

retinopathy at an early stage.

ii

ACKNOWLEDGEMENTS

First of all, I would like to express my gratitude to my Director of Studies, Dr. Vasile

Palade, for his time, guidance and support for me during my studies. I am also grateful

to my supervisors, Dr. James Shuttleworth and Prof. Chrisina Jayne. I would like to

thank them for being a great supervisory team.

My gratitude also goes to my previous supervisory team, Dr. Rajeev Bali and Prof.

Raouf Naguib. Under their supervision, I learnt a lot from their vast knowledge and

experience.

The deepest gratitude and thanks go to the Universiti Teknikal Malaysia Melaka and

Ministry of Higher Education Malaysia for sponsoring this PhD research. I am also

thankful to the Ministry of Health Malaysia and the Melaka Hospital, Malaysia for

providing access and approve the research and data collection.

Special thanks go to Dr. Raja Norliza Raja Omar and the staff at the Eye Clinic,

Department of Ophthalmology, Melaka Hospital, Malaysia for the medical guidance,

supports and for making my research data collection possible.

Thanks go to my wonderful parents, my siblings, other members of my family and my

friends for their countless support.

Finally, I would like to thank everyone who was involved directly or indirectly

throughout this research.

iii

TABLE OF CONTENTS

ABSTRACT ...................................................................................................................... i

ACKNOWLEDGEMENTS ............................................................................................. ii

LIST OF FIGURES ......................................................................................................... vi

LIST OF FIGURES FOR APPENDICES ..................................................................... viii

LIST OF TABLES .......................................................................................................... ix

LIST OF TABLES FOR APPENDICES ......................................................................... x

LIST OF EQUATIONS .................................................................................................... x

LIST OF ABBREVIATIONS ......................................................................................... xi

LIST OF PUBLISHED PAPERS ON THIS RESEARCH ............................................ xii

1 INTRODUCTION ......................................................................................................... 1

1.1 Diabetic Retinopathy Screening and the Challenges Faced by the Malaysian

Government .................................................................................................................. 2

1.2 Research Aims and Objectives ............................................................................... 5

1.3 Motivation and Contributions of the Thesis ........................................................... 8

1.4 Research Methodology ......................................................................................... 11

1.5 Thesis Overview ................................................................................................... 13

2 BACKGROUND AND LITERATURE REVIEW ..................................................... 15

2.1 Diabetes Mellitus .................................................................................................. 15

2.2 Diabetic Retinopathy ............................................................................................ 17

2.3 Epidemiology of Diabetic Retinopathy ................................................................ 21

2.3.1 Diabetic Retinopathy and Global Epidemiology ........................................... 21

2.3.2 Diabetic Retinopathy Epidemiology in Malaysia .......................................... 23

2.4 Classification of Diabetic Retinopathy ................................................................. 26

2.5 Diabetic Retinopathy Screening ........................................................................... 27

2.5.1 Diabetic Retinopathy Grading ....................................................................... 33

2.5.2 Examination Schedule ................................................................................... 38

2.5.3 Diabetic Retinopathy Management and Treatment ....................................... 40

2.6 Diabetic Retinopathy Image Processing ............................................................... 41

2.6.1 Image Preprocessing ...................................................................................... 41

2.6.2 Feature Extraction ......................................................................................... 47

2.6.3 Classification ................................................................................................. 51

2.6.3.1 Support Vector Machines ....................................................................... 52

2.6.3.2 Neural Networks ..................................................................................... 53

2.6.3.3 Other Classification Methods ................................................................. 55

2.7 Fuzzy Image Processing ....................................................................................... 58

2.8 Summary ............................................................................................................... 62

3 RESEARCH METHODOLOGY AND DATA COLLECTION ................................ 63

3.1 Research Design ................................................................................................... 63

3.2 Study Population................................................................................................... 65

3.3 Location of the Research Study ............................................................................ 65

iv

3.4 Data Collection ..................................................................................................... 65

3.4.1 Primary Data .................................................................................................. 66

3.4.2 Secondary Data .............................................................................................. 66

3.5 Collecting Data Procedure .................................................................................... 67

3.5.1 Collecting Primary Data through Observation .............................................. 67

3.5.2 Collecting Primary Data through Interview .................................................. 69

3.5.3 Collecting Secondary Data through Existing Fundus Images ....................... 69

3.6 Data Management and Analysis Procedure .......................................................... 70

3.7 Ethical Considerations .......................................................................................... 71

3.8 Benchmark Public Data Sets ................................................................................ 72

3.9 Developed Data Set .............................................................................................. 74

3.10 Summary ............................................................................................................. 79

4 CONTRIBUTIONS TO DEVELOPING SYSTEMS FOR DIABETIC

RETINOPATHY AND MACULOPATHY DETECTION ........................................... 81

4.1 Basic System for General Diabetic Retinopathy Detection (System I) ................ 82

4.1.1 Image Preprocessing ...................................................................................... 84

4.1.2 Feature Extraction ......................................................................................... 87

4.1.3 Classification ................................................................................................. 88

4.1.4 System Results and Evaluation ..................................................................... 89

4.2 Microaneurysms Detection ................................................................................... 91

4.2.1 Automatic Detection of Microaneurysms in Colour Fundus Images

Using Vessel Segmentation and Features Extraction (System II) ......................... 91

4.2.1.1 Image Preprocessing ............................................................................... 93

4.2.1.2 Feature Extraction................................................................................... 94

4.2.1.3 Classification .......................................................................................... 95

4.2.1.4 System Results and Evaluation............................................................... 95


Using a Combination of Image Preprocessing Techniques and Circular Hough

Transform (System III) ........................................................................................... 97

4.2.2.1 Image Preprocessing ............................................................................... 98

4.2.2.2 Circular Hough Transform ..................................................................... 98

4.2.2.3 System Results and Evaluation............................................................. 100


Using Fuzzy Image Processing (System IV) ........................................................ 102

4.2.3.1 Image Preprocessing ............................................................................. 103

4.2.3.2 Microaneurysms Detection ................................................................... 105

4.2.3.3 System Results ...................................................................................... 105

4.2.4 Automatic Detection of Microaneurysms Using Fuzzy Histogram

Equalisation, Fuzzy Filtering and Fuzzy Edge Detection (System V) ................. 107


4.2.4.2 Localisation and Detection of Microaneurysms ................................... 111

4.2.4.3 Classification ........................................................................................ 111

v


4.3 Maculopathy and Diabetic Retinopathy Detection ............................................. 114

4.3.1 Detection of Diabetic Retinopathy and Maculopathy in Eye Fundus

Images Using Fuzzy Image Processing (System VI) ........................................... 116


4.3.1.2 Feature Extraction................................................................................. 119

4.3.1.3 Classification ........................................................................................ 120


4.3.2 Detection of Diabetic Retinopathy and Maculopathy in Eye Fundus

Images Using Fuzzy Image Processing and Retinal Structures Segmentation

(System VII) ......................................................................................................... 122


4.3.2.2 Retinal Structure Extraction ................................................................. 125

4.3.2.3 Exudate and Maculopathy Detection .................................................... 129

4.3.2.4 Features Extraction ............................................................................... 130

4.3.2.5 Classification ........................................................................................ 131


4.4 Summary ............................................................................................................. 132

5 OVERALL RESULTS ANALYSIS AND DISCUSSION ....................................... 135

5.1 Developed Data Set Analysis ............................................................................. 135

5.2 System Results Analysis ..................................................................................... 140

5.2.1 Confusion Matrix ......................................................................................... 140

5.2.2 Statistical Analysis ...................................................................................... 141

5.3 Discussion ........................................................................................................... 148

5.4 Summary ............................................................................................................. 151

6 CONCLUSIONS AND FUTURE WORK ................................................................ 152

6.1 Review of the Work and Contributions .............................................................. 152

6.2 Future Work ........................................................................................................ 155

6.3 Summary ............................................................................................................. 158

REFERENCES ............................................................................................................. 159

APPENDICES .............................................................................................................. 171

vi

LIST OF FIGURES

Figure 1.1 Methodology of the research ......................................................................... 12

Figure 2.1 Retina (eyeSmart, 2014) ............................................................................... 17

Figure 2.2 Diagrammatic representation of the different types of damage to capillaries

in diabetic retinopathy (Taylor and Batey, 2012:24) (a) Normal capillary (b)

Microaneurysm (c) Blot haemorrhage (d) Exudate ................................................ 18

Figure 2.3 Features of diabetic retinopathy (Ministry of Health Diabetic Retinopathy

Screening Team, 2012b) ......................................................................................... 20

Figure 2.4 Normal retina compared to retina with diabetic retinopathy signs ............... 21

Figure 2.5 Global causes of visual impairment, including blindness, as percentage

(World Health Organization, 2012b:6) ................................................................... 23

Figure 2.6 Global causes of blindness as a percentage of global blindness in 2010

(World Health Organization, 2012b:6) ................................................................... 23

Figure 2.7 Percentage distribution of diabetic retinopathy cases, glaucoma or lens-

induced glaucoma, 2002-2013 (NED Steering Committee Members, 2015:112) .. 25

Figure 2.8 Diabetic retinopathy screening tools (Ministry of Health Diabetic

Retinopathy Screening Team, 2012c:14) ................................................................ 29

Figure 2.9 Screening process of diabetic retinopathy to prevent blindness (Ministry of

Health Malaysia, Malaysian Society of Ophthalmology and Academy of Medicine

of Malaysia, 2011) .................................................................................................. 31

Figure 2.10 Fundus camera photography (Ministry of Health Diabetic Retinopathy

Screening Team, 2012b, 2012d) ............................................................................. 32

Figure 2.11 Good quality fundus photos (Ministry of Health Diabetic Retinopathy

Screening Team, 2012b) ......................................................................................... 32

Figure 2.12 Normal fundus image (Ministry of Health Diabetic Retinopathy Screening

Team, 2012b) .......................................................................................................... 33

Figure 2.13 Basic two-class classifications with Support Vector Machines (The

Mathworks, Inc., 2016a) ......................................................................................... 53

Figure 2.14 Brightness Preserving Dynamic Fuzzy Histogram Equalization stages ..... 59

Figure 2.15 Membership functions of the inputs and outputs (The Mathworks, Inc.,

2016c) ..................................................................................................................... 61

Figure 2.16 Edge detection using fuzzy logic output (The Mathworks, Inc., 2016c) .... 61

Figure 3.1 Research development design ....................................................................... 64

Figure 3.2 Research study location................................................................................. 66

vii

Figure 3.3 Fundus images capturing and the grading process ........................................ 68

Figure 3.4 Examples of images in the dataset ................................................................ 75

Figure 3.5 Expert diagnosis file ...................................................................................... 76

Figure 3.6 Expert diagnosis summary ............................................................................ 76

Figure 3.7 Expert diagnosis average............................................................................... 77

Figure 3.8 Main web page for the developed dataset ..................................................... 78

Figure 4.1 Block diagram of the proposed general automatic screening and classification

of diabetic retinopathy ............................................................................................ 83

Figure 4.2 Image preprocessing...................................................................................... 87

Figure 4.3 Snapshot of the proposed system user interface ........................................... 90

Figure 4.4 Block diagram of the proposed automatic detection of microaneurysms ..... 92

Figure 4.5 Preprocessing the output image .................................................................... 94

Figure 4.6 Snapshot of the proposed system user interface ........................................... 96

Figure 4.7 Block diagram of the proposed automatic detection of microaneurysms using

Circular Hough Transform ...................................................................................... 98

Figure 4.8 Microaneurysms detection by using the Circular Hough Transform ............ 99

Figure 4.9 Snapshot of the user interface of System III ............................................... 100

Figure 4.10 Block diagram of System IV for the automatic detection of microaneurysms

............................................................................................................................... 104

Figure 4.11 Preprocessing the output image ................................................................ 104

Figure 4.12 Snapshot of the user interface of System IV ............................................. 106

Figure 4.13 Block diagram of System V for the automatic detection of microaneurysms

............................................................................................................................... 108

Figure 4.14 Membership functions for inputs and outputs ........................................... 110

Figure 4.15 Preprocessing the output image with fuzzy approaches ........................... 111

Figure 4.16 Snapshot of the proposed system user interface ....................................... 113

Figure 4.17 Block diagram of System VI for the automatic detection of diabetic

retinopathy and maculopathy using fuzzy image processing................................ 117

Figure 4.18 Membership functions for inputs and outputs ........................................... 118

Figure 4.19 Preprocessing of the output image ............................................................ 119


viii

Figure 4.21 Block diagram of System VII for the automatic detection of diabetic

retinopathy and maculopathy using fuzzy image processing................................ 123

Figure 4.22 Preprocessing the output image ................................................................ 125

Figure 4.23 Extraction of retinal structures .................................................................. 128

Figure 4.24 Exudates and maculopathy extraction output image ................................. 130


Figure 5.1 Boxplot assessment ..................................................................................... 137

Figure 5.2 Histogram assessment ................................................................................. 138

Figure 5.3 System development workflow ................................................................... 147

LIST OF FIGURES FOR APPENDICES

Figure A.1 Data collection and diabetic retinopathy screening process ....................... 171

Figure A.2 No diabetic retinopathy images .................................................................. 172

Figure A.3 Mild DR without maculopathy images ...................................................... 172

Figure A.4 Mild DR with maculopathy images ........................................................... 173

Figure A.5 Moderate DR without maculopathy images ............................................... 173

Figure A.6 Moderate DR with maculopathy images .................................................... 174

Figure A.7 Severe DR without maculopathy images ................................................... 174

Figure A.8 Severe DR with maculopathy images ........................................................ 175

Figure A.9 Proliferative DR without maculopathy images .......................................... 175

Figure A.10 Proliferative DR with maculopathy images ............................................. 176

Figure A.11 Advanced diabetic eye disease images .................................................... 176

Figure B.1 ‘Home’ page ............................................................................................... 184

Figure B.2 ‘About’ page ............................................................................................... 184

Figure B.3 ‘Team’ page ................................................................................................ 185

Figure B.4 ‘Publications’ page ..................................................................................... 185

Figure B.5 ‘Dataset’ page and ‘Download’ link .......................................................... 186

Figure B.6 ‘Contact’ page ............................................................................................ 186

ix

LIST OF TABLES

Table 2.1 Terminology definition of diabetic retinopathy signs (adapted from Taylor

and Batey, 2012) ..................................................................................................... 19

Table 2.2 Sensitivity and specificity of diabetic retinopathy tools (Ministry of Health

Malaysia, Malaysia Society of Ophthalmology and Academy of Medicine of

Malaysia, 2011) ....................................................................................................... 29

Table 2.3 Retinopathy stages and findings (Ministry of Health Diabetic Retinopathy

Screening Team, 2012b) ......................................................................................... 34

Table 2.4 International Clinical Diabetic Retinopathy Disease Severity Scale

(Wilkinson et al., 2003) .......................................................................................... 37

Table 2.5 Diabetic Macular Edema Disease Severity Scale (Wilkinson et al., 2003).... 38

Table 2.6 Timing of the first screening (Ministry of Health Malaysia, Malaysian Society

of Ophthalmology and Academy of Medicine of Malaysia, 2011) ........................ 39

Table 2.7 Recommended follow-up schedule (American Academy of Ophthalmology

Retina Panel, 2008) ................................................................................................. 39

Table 2.8 Criteria for urgent referral (National Institute for Clinical Excellence, 2002) 40

Table 2.9 Summary of treatment for diabetic retinopathy (Ministry of Health Malaysia,

Malaysian Society of Ophthalmology and Academy of Medicine of Malaysia,

2011) ....................................................................................................................... 40

Table 2.10 Summary of preprocessing techniques used in diabetic retinopathy screening

research ................................................................................................................... 43

Table 2.11 Summary of feature extraction methods used in previous diabetic

retinopathy screening research ................................................................................ 48

Table 2.12 Summary of machine learning techniques for classification used in previous

diabetic retinopathy screening research .................................................................. 56

Table 3.1 Sensitivity, specificity and accuracy formula ................................................. 71

Table 3.2 Expert diagnosis summary ............................................................................. 77

Table 3.3 Summary of datasets used for systems development ..................................... 80

Table 4.1 Feature extraction in the proposed system ..................................................... 88

Table 4.2 Average results when using the four classifiers ............................................. 90

Table 4.3 Feature extraction in the proposed microaneurysms detection system .......... 95

Table 4.4 Average results when using the four classifiers ............................................. 96

Table 4.5 T-test analysis ............................................................................................... 101

Table 4.6 Chi-square analysis ....................................................................................... 101

x

Table 4.7 Summary results for Systems IV(a) and IV(b) ............................................. 106

Table 4.8 Summary results for System V(a), V(b), V(c) and V(d) .............................. 113

Table 4.9 Summary results for System VI ................................................................... 121

Table 4.10 Average results when using the four classifiers ......................................... 132

Table 5.1 Expert diagnosis summary categorisation .................................................... 136

Table 5.2 ANOVA multiple comparisons .................................................................... 139

Table 5.3 Chi-square analysis ....................................................................................... 140

Table 5.4 Confusion matrix analysis of the proposed systems..................................... 142

Table 5.5 Inferential statistical analysis of the proposed systems ................................ 145

LIST OF TABLES FOR APPENDICES

Table A.1 Data collection methods .............................................................................. 177

Table A.2 Data collection assessment .......................................................................... 178

Table A.3 Expert diagnosis .......................................................................................... 179

Table A.4 Expert diagnosis summary .......................................................................... 180

Table A.5 Findings summary ....................................................................................... 182

Table A.6 Discussion summary .................................................................................... 183

LIST OF EQUATIONS

(4.1) ................................................................................................................................ 85

(4.2) ................................................................................................................................ 86

xi

LIST OF ABBREVIATIONS

DM Diabetes Mellitus

DR Diabetic Retinopathy

DME Diabetic Macula Edema

MOHM Ministry of Health Malaysia

NPDR Non-proliferative Diabetic Retinopathy

PDR Proliferative Diabetic Retinopathy

WHO World Health Organization

NHMS National Health and Morbidity Survey

T1DM Type 1 Diabetes Mellitus

T2DM Type 2 Diabetes Mellitus

IRMAs Intraretinal Microvascular Anomalies

NED National Eye Database

SVM Support Vector Machines

HOS Higher Order Spectra

DWT Discrete Wavelet Transform

ANN Artificial Neural Network

D-FNN Dynamic Fuzzy Neural Network

RBFNN Radial Basis Function Neural Network

PNN Probabilistic Neural Network

kNN k-Nearest Neighbours

CNN Convolutional Neural Networks

xii

LIST OF PUBLISHED PAPERS ON THIS RESEARCH

Journal:

Rahim, S, S., Jayne, C., Palade, V., and Shuttleworth, J. (2015) ‘Automatic Detection

of Microaneurysms in Colour Fundus Images for Diabetic Retinopathy

Screening’. Journal of Neural Computing and Applications 521, 1-16

Rahim, S. S., Palade, V., Shuttleworth, J., and Jayne, C. (2016) ‘Automatic Screening

and Classification of Diabetic Retinopathy and Maculopathy Using Fuzzy Image

Processing’. in Zhong, N., and Peng, H. (ed.) Brain Informatics 40708, 1-19

Conference Papers:

Rahim, S. S., Palade, V., Shuttleworth, J., and Jayne, C. (2014) ‘Automatic Screening

and Classification of Diabetic Retinopathy Fundus Images’. in Mladenov, V. et

al. (ed.) Proceedings of 15th

International Conference on Engineering

Applications of Neural Networks, EANN 2014, Communications in Computer

and Information Science 459. held 5-7 September 2014 at Sofia, Bulgaria.

Switzerland: Springer, 113-122

Rahim, S. S., Palade, V., Shuttleworth, J., Jayne, C., and Raja Omar, R. N. (2015)

‘Automatic Detection of Microaneurysms for Diabetic Retinopathy Screening

Using Fuzzy Image Processing’. in Iliadis, L., and Jayne, C. (ed.) Proceedings of

16th

International Conference on Engineering Applications of Neural Networks,

EANN 2015, Communications in Computer and Information Science 517. held

25-28 September 2015 at Rhodes, Greece. Switzerland: Springer, 69-79

Rahim, S. S., Palade, V., Jayne, C., Holzinger, A., and Shuttleworth, J. (2015)

‘Detection of Diabetic Retinopathy and Maculopathy in Eye Fundus Images

Using Fuzzy Image Processing’. in Guo, Y. et al. (ed.) Proceedings of 8th

International Conference on Brain Informatics and Health, BIH 2015, LNAI

9250. held 30 August-2 September 2015 at London, UK. Switzerland: Springer,

379-388

1

1 INTRODUCTION

Diabetic Retinopathy (DR) is one of the diabetes complications and is an important

cause of visual disability and blindness. It is vital to have a regular eye examination for

initial detection and early treatment. This thesis is about the development of a medical

decision support system for automatic diabetic retinopathy screening and classification

in eye fundus images.

The scope of this research focuses on the Asian country of Malaysia. The research was

sponsored by the Ministry of Higher Education in Malaysia and the Universiti Teknikal

Malaysia Melaka (UTeM). In addition, the research also received support, particularly

in the contribution of expertise, from the Department of Ophthalmology at the Melaka

Hospital, Malaysia, under the Ministry of Health for Malaysia.

In order to develop an accurate and efficient diabetic retinopathy screening system, the

capabilities of image processing techniques are investigated in this research work.

Several highly efficient image processing techniques are implemented and tested to

evaluate the system performance. The main image processing techniques used in this

research involve fuzzy image processing on eye fundus images, which has not been

previously investigated. A more reliable screening system can be produced with fuzzy

image processing capability and therefore enable the achievement of the screening

general aim, which is the earlier detection of sight threatening problems to ensure

prompt treatment for the prevention of vision loss.

This chapter introduces diabetic retinopathy and its screening. It focuses on Malaysia

and highlights some challenges faced by the Malaysian healthcare system in dealing

with diabetic retinopathy screening. Section 1.2 presents the research aims and

objectives. Section 1.3 explains the research motivation, alongside the contributions of

the thesis. Finally, Section 1.4 presents the overview of the chapters of this thesis.

2

1.1 Diabetic Retinopathy Screening and the Challenges Faced

by the Malaysian Government Diabetic Retinopathy (DR) is a diabetes mellitus complication, which also include

stroke, cardiovascular disease, diabetic nephropathy and diabetic neuropathy. The

retinal capillaries damage occurs in diabetes mellitus. Diabetic retinopathy can be

visualised only in the retina (Taylor and Batey, 2012), which is a layer of tissue.

Diabetic retinopathy happens through lasting of small blood vessels damage in the

retina, which eventually results in blindness. Hence, an effective screening of diabetic

retinopathy is important for early treatment, as well as an effective management of risk

factor to prevent diabetic complications.

The prevalence of diabetes globally presented by the World Health Organization

(WHO), reported in 2014 was estimated to be 9% among adults aged 18 and above

(2012a). Diabetes contributes to about one percent of global blindness (2012b).

Globally, 4.8% of the 37 million blindness cases is caused by the diabetic retinopathy

and approximately 366 million will be affected by diabetes mellitus worldwide in the

year 2030 (World Health Organization, 2005). In addition, diabetes has also been

predicted to be the seventh leading cause of death by the year 2030 (Mathers and

Loncar, 2006). The initial National Health and Morbidity Survey I (NHMS I) for

Malaysia was conducted in 1986. The survey reveals a prevalence of diabetes mellitus is

6.3%. In the year 1996, the percentage in NHMS II had increased to 8.3% and, the

newest NHMS III 2006 report shows that the total diabetes mellitus prevalence is 14.9%

(Letchuman et al., 2010).

Diabetes Mellitus (DM) is a complex disease resulting in severe complications in

various parts of the body. Nevertheless, good control of DM will avoid or delay various

complications, including diabetic retinopathy. The length of DM is associated with the

diabetic retinopathy incidence, and it has been reported that more than 75% of diabetes

patients have some diabetic retinopathy form after 20 years of the disease (World

Health Organization, 2005). Thus, screening and early treatment can avoid significant

loss of vision. Such efforts to control this enduring disease as well as the early

complications detection such as diabetic retinopathy should be strengthened, because

diabetic retinopathy is an asymptomatic condition in its initial stage (Taylor and Batey,

3

2012; Ministry of Health Malaysia, Malaysian Society of Ophthalmology and Academy

of Medicine Malaysia, 2011). It is also stated that diabetics are twenty-five times more

likely to develop blindness compared to the general population (Health Technology

Assessment Unit, Medical Development Division, Ministry of Health Malaysia, 2002).

Considering these complications and the rising numbers of diabetic patients in

Malaysia, a programme of diabetic retinopathy screening in the country must be

complete, covering all people with DM in Malaysia. To achieve this, significant

resources will be required for the management of the condition including human

resources, to increase the current workload within the field of disease diagnostics.

Diabetic retinopathy can only be identified through medical eye examination since it is

asymptomatic in its initial stage. The diabetes management at the facilities of the

Ministry of Health Malaysia (MOHM) is presently being performed in health clinics, in

addition in polyclinics, specialised clinics as well as hospital wards. In Malaysia,

screening is currently performed by general practitioners, clinicians in hospital based

diabetes centres, ophthalmologists, optometrists or a technician and a medically trained

photographic interpreter, in the case of photography (Health Technology Assessment

Unit, Medical Development Division, Ministry of Health Malaysia, 2002). There are

many modalities of screening available to detect and classify diabetic retinopathy. One

of the most common techniques used is ophthalmoscopy. Non-mydriatic digital fundus

photography is also popular (Ministry of Health Malaysia, Malaysian Society of

Ophthalmology and Academy of Medicine Malaysia, 2011).

Different screening modalities performed by different practitioners will produce a wide

variation of sensitivity and specificity. The screening tools include the following: the

direct and indirect ophthalmoscope, the slit lamp biomicroscope, the mydriatic fundus

camera and the non-mydriatic fundus camera. The non-mydratic fundus camera has

high sensitivity and specificity among other advantages. For example pupillary

dilatation is not required, especially if the room is adequately darkened, promoting

compliance, efficiency and safety (Ministry of Health Malaysia, Malaysian Society of

Ophthalmology and Academy of Medicine Malaysia, 2011). Trained primary care

clinicians are needed for screening of diabetic retinopathy to increase the accuracy of

4

interpretation and grading. Proper training among all healthcare personnel therefore is

essential. Specialised personnel for retinal screening and grading need specific training

and regular performance assessment. Moreover, the Clinical Practice Guidelines

Screening for Diabetic Retinopathy requires that the module of training should comprise

clinical skills and knowledge, computer imaging and skills, in addition operational

concerns and training of fundus grading (Ministry of Health Malaysia, Malaysian

Society of Ophthalmology and Academy of Medicine Malaysia, 2011).

Diabetes Mellitus is a growing problem among increasing numbers of diabetics every

year. Subsequently, there are several challenges faced by the Ministry of Health

Malaysia in diabetic retinopathy handling cases (Ministry of Health Malaysia, 2012a),

including:

i. Inadequate diabetic eye screening programs

In order to perform successful eye screening, a team of trained healthcare

personnel is required. Fewer screening teams, especially in rural hospitals,

have significantly decreased the number of the eye screening programs. At

the moment, screening programs are exclusively performed at primary health

care centres (selected health clinic with fundus camera), hospital or clinics

with eye care providers such as ophthalmology and optometry clinics.

According to a relatively recent report on diabetic retinopathy screening by

the Unit of Health Technology Assessment, only 24 out of the 114 Ministry

of Health hospitals have a department of ophthalmology, while

ophthalmologists aim to visits other hospitals regularly (Health Technology

Assessment Unit, Medical Development Division, Ministry of Health

Malaysia, 2002).

ii. Inadequate resources to complete the task

The main resources needed for the screening are trained staff and fundus

cameras. All healthcare personnel involved in screening require proper

training before they can join the programs. There is a need for training on

5

how to screen the images and how to improve the accuracy of interpretation

and grading, in addition in terms of sensitivity and specificity.

The screening tools and techniques used in the program are the other

important factors to be considered. There are many available screening

modalities used for diabetic retinopathy screening. Ophthalmoscopy is a

popular screening method, but non-mydriatic digital camera is also widely

prescribed due to its high sensitivity and specificity. The fundus camera

however is limited throughout Malaysian hospitals and health clinics. In the

year of 2011, the total number of fundus camera available at health clinics

under the Ministry of Health was only 107 (Ministry of Health Malaysia,

Malaysian Society of Ophthalmology, Academy of Medicine Malaysia,

2011). This lack of vital screening resources will invariably result in longer

waiting lists for initial screenings and referrals to ophthalmologist. This will

ultimately lead to more serious eye complications.

iii. Poor patient information or awareness

One of the barriers in handling diabetic retinopathy is the patient factor. A

lack of awareness of the possible eye complications from diabetes mellitus is

one of the factors that have decreased the frequency of diabetic retinopathy

screening. Moreover, other factors which impede the level of diabetic

retinopathy awareness amongst Malaysians are eye care services poor access

and dissimilar cultural beliefs. As such, patients should be aware that regular

eye examinations are important.

1.2 Research Aims and Objectives The aim of this research is to investigate automatic methods for diabetic retinopathy

detection that can contribute towards improving diabetic retinopathy management and,

subsequently, to develop an efficient system for diabetic retinopathy screening.

Basically, the proposed diabetic retinopathy research consists of three types of systems.

Firstly, the thesis will present the development of a basic system for the screening and

classification of diabetic retinopathy using eye fundus images, which is a system for

6

general detection for diabetic retinopathy screening and will classify images into two

respective cases: Normal and Diabetic Retinopathy. The research will then focus on the

microaneurysms detection which are the earliest diabetic retinopathy signs.

Different image processing techniques, including fuzzy image processing, are

implemented in a variety of detection systems for microaneurysms which classify

images into two main categories. The first categorisation classifies them into detected

and non-detected cases. The second categorisation is based on Normal (No DR) and

Diabetic Retinopathy cases. In addition, the thesis presents the fuzzy-based image

processing decision support systems for diabetic retinopathy and maculopathy detection

in eye fundus images. The proposed systems classify the images into two types of

classification, in order to generate a diversity of results and system performance

analysis, which are the two above cases (Normal and Diabetic Retinopathy) and an

additional ten cases which follow the ophthalmologists’ practice and provide more

details. The second classification involves No Diabetic Retinopathy and the other nine

detailed classes of the DR cases: Mild DR without maculopathy, Mild DR with

maculopathy, Moderate DR without maculopathy, Moderate DR with maculopathy,

Severe DR without maculopathy, Severe DR with maculopathy, Proliferative DR

without maculopathy, Proliferative DR with maculopathy and Advanced Diabetic Eye

Disease (ADED).

In order to assist screeners to classify the retinal images effectively and with high

confidence, an accurate retinal screening system is necessary. Therefore, to develop a

diabetic retinopathy screening grading and classification system, effective techniques of

image processing must be used. This research project examines the use of the fundus

images for detecting the diabetic retinopathy features presence in the eyes. This is a

particularly challenging problem and this thesis proposes novel use of image processing

techniques in order to automatically detect the stages of retinopathy. To achieve this

aim, highly efficient and accurate image processing techniques must be used to produce

an effective screening of diabetic retinopathy.

7

Despite the existence of a range of image processing techniques, the need for highly

effective and specialised image processing techniques in this case cannot be over

emphasised. Factors such as the fundus images suffering from noise and latency are

often encountered, necessitating calibration and filtering before the images can be used

reliably. In addition, the quality of the image depends on the skills applied by the

paramedic in capturing the eye fundus images, as well as on other factors including the

quality of the equipment and possible distractions from the environment. Due to these

facts, all healthcare staff requires proper training before they are qualified and equipped

for diabetic retinopathy screening. This is important as it can help increase the

likelihood of accurate interpretation and grading. In addition to the lengthy and rigorous

training of healthcare personnel before qualification, a growing challenge faced by the

healthcare sector is the fact that diabetes mellitus is on the increase, with higher

numbers of diabetics each year. It has also been highlighted by governments and other

relevant stakeholders that the diabetic eye screening program is inadequate, as are the

resources to complete the task, in addition to poor patient information or awareness

(Ministry of Health Diabetic Retinopathy Screening Team, 2012b).

Further to the range of complications associated with images captured by the fundus

camera, there is also the need for an experienced paramedic to diagnose whether the

patient has any conditions (i.e., diabetic retinopathy). This screening phase is carried out

manually by the paramedic who looks for any changes (abnormalities) on the retinal

image, making the whole diagnostic process highly convoluted and protracted.

Based on the aforementioned reasons, in order to pursue this study primary research

outcome, which is the development of a computer-based imaging tool, a method must

be created in order to effectively detect important features on the fundus images and

efficiently classify patients into the correct retinopathy stages. This automatic diabetic

retinopathy grading will facilitate a reduction in the burden of manual grading for the

screening team, and help alleviate the pressure on the limited eye screening centres in

Malaysia (Ministry of Health Malaysia, Malaysian Society of Ophthalmology and

Academy of Medicine Malaysia, 2011). As a result of early detection, it would also help

ophthalmologists to treat patients before their conditions worsen and, most importantly,

8

increase the chance of protecting the patient’s vision. Moreover, an automatic diabetic

retinopathy system would diagnose it in a faster and more efficient way. In addition, as

suggested by the available literature, the initial detection of retinopathy, the existing

retinopathy monitoring with consistent fundus examinations and effective laser

treatment at suitable times, are among the key measures to prevent visual loss from

diabetic retinopathy (Health Technology Assessment Unit, Medical Development

Division, Ministry of Health Malaysia, 2002).

The main objectives of the research described in this thesis are as follows:

i. To develop an automatic screening and classification systems for diabetic

retinopathy using fundus images in order to detect diabetic retinopathy at an

early stage.

ii. To propose novel use of image processing and machine learning techniques

for early detection of the signs of diabetic retinopathy.

The research introduced novel use of image processing techniques for the automated

detection of retinopathy stages, including the combination of various pre-processing

techniques as well as fuzzy image processing techniques, such as fuzzy histogram

equalisation, fuzzy filtering and fuzzy edge detection. In addition, the research proposed

the use of Circular Hough Transform and various machine learning classifiers.

1.3 Motivation and Contributions of the Thesis

Eye screening is important for the early detection and treatment of diabetic retinopathy.

Regular screening can help detect patients with diabetes at an early stage thus, earlier

identification of any retinopathy can allow changes in blood pressure or blood glucose

to be managed efficiently to slow the rate of progression of the disease. The importance

of the proposed research is to overcome the current problems faced in the diabetic

retinopathy screening process, such as:

9

i. Manual diagnosis by the ophthalmologist

Currently, clinicians use non-mydratic fundus cameras to capture retinal

images. Based on the image produced from the fundus camera, the

experienced screening team will diagnose whether or not patients have any

conditions (including diabetic retinopathy). The diagnosis is carried out

manually by screeners who assess any changes (abnormalities) on the retinal

image. This process is both laborious and prone to error. Therefore, a

computer-based imaging tool is needed to effectively detect the signs of

diabetic retinopathy, allowing ophthalmologists to gain a suitable window in

which to treat patients, before serious damage occurs, thus increasing the

chance of protecting the patient’s vision. It will also help decrease the

workload for healthcare personnel in the diabetic retinopathy screening

process.

ii. Time taken and limitations of screening resources

The proposed automatic diabetic retinopathy system would help save time,

costs and ultimately the vision of patients. With appropriate automation (i.e.,

decision support systems) in place, preventative actions to protect vision can

then be taken earlier and therefore can help reduce the number of diabetic

retinopathy problems, in addition to the risk of blindness. A decision support

system for clinical diagnosis would contribute greatly in assisting with the

management and detection of diabetic retinopathy. An automatic system will

assist an ophthalmologist (or optometrist) to detect diabetic retinopathy (and

its detailed classification) in a more efficient and faster way compared with

manual analysis, which is more time-consuming. As a result, the proposed

system will indirectly assist in the process of recommended follow-up

schedules for each category of diabetic retinopathy based on the system

detection. Furthermore, the development of the proposed system will

contribute to overcoming the diabetic retinopathy screening limitations

inherent in the present manual screening procedure, especially given the

problems of inadequately trained staff and the use of the fundus camera, as

highlighted in Section 1.1.

10

iii. Developing effective techniques of image processing for the diabetic

retinopathy detection

Diabetic retinopathy screening is a popular research area and many

researchers focus on and contribute to the advancement of this study area.

Most researchers focus on finding and proposing an accurate technique or

method for detecting certain features of diabetic retinopathy through

exploring the eye fundus images. Although there have been immense

advancements in this area of research, there are still lacunae or spaces for

improvement. The proposed techniques in this research will most notably

benefit the realm of image processing in a number of areas or ways that

include the provision of an accurate method for effectively detecting features

of diabetic retinopathy.

Based on the general objectives in Section 1.2 above, in particular the

second objective, the highlighted contributions of this thesis include:

implementing image processing techniques combination for the general

diabetic retinopathy screening detection (Rahim et al., 2014), investigating

image processing techniques combination for the diabetic retinopathy

features detection, focusing on microaneurysms, an important early feature

of diabetic retinopathy (Rahim et al., 2015a; 2015b) and the evaluation of

image processing techniques combination for the diabetic retinopathy and

diabetic maculopathy detection (Rahim et al., 2015c; 2016). In addition, the

contributions of this thesis include employing the novel use of fuzzy image

processing techniques for the pre-processing stage of medical images, i.e.,

eye fundus images for diabetic retinopathy screening (Rahim et al., 2015a;

2015b; 2015c; 2016), as well as implementing a new online dataset

containing normal and diabetic retinopathy fundus images (Rahim et al.,

2015b; 2015c; 2016) and finally, testing a new method for macula region

localisation in order to detect maculopathy (Rahim et al., 2016).

To summarise, diabetes mellitus is a main health problem. One of the diabetes mellitus

health effects is diabetic retinopathy, which causes blindness. Therefore, an effective

11

tool to help in the diabetic retinopathy detection is essential. A computer-based imaging

screening method is needed to be developed where effective and cost-effective

approaches are required. Automatic detection systems of diabetic retinopathy for

patients with diabetes using the eye fundus photography will help the screening process

by providing a user- or patient-friendly approach in addition to a cost-effective

screening tool. Automatic classification systems with a high accuracy of diabetic

retinopathy screening will help in decreasing the workload for healthcare personnel in

the process of the early detection of diabetic retinopathy. It would also be helpful to

patients in terms of early treatment, which could prevent or ameliorate substantial visual

loss. This thesis proposes an automatic diabetic retinopathy detection system, and also

introduces a new dataset of fundus images, which would be beneficial to retinal imaging

researchers and practitioners, especially in the diabetic retinopathy screening field.

1.4 Research Methodology The research methodology is concerned with the base of the inverted triangle in Figure

1.1. The figure shows that the main components of this research are diabetic retinopathy

screening, fundus images as input and image processing techniques. The developed

system consists of three stages: image pre-processing, feature extraction and

classification. In order to validate the systems output performance, the results generated

from the developed system are compared to the expert findings and several analyses are

performed.

Understanding the diabetic retinopathy screening process, including the diabetic

retinopathy development and diabetic retinopathy signs, are essential and information is

collected through observation and interview techniques. The fundus images, which are

the main data for this research, are extracted from the personal computer that is attached

to the fundus camera. Each patient folder consists of the patient information and the eye

fundus images. The automatic diabetic retinopathy detection systems are developed by

employing a novel use of image processing techniques. The developed systems consists

of three types of system development, which are the general detection of diabetic

retinopathy, the development of an automatic system for microaneurysms detection and

the development of an automatic diabetic retinopathy and maculopathy detection

12

system. The systems are evaluated with the combination of normal and diabetic

retinopathy fundus images from a new data set collected during this research and also

from several public datasets available as benchmark data. A thorough system

performance analysis has been undertaken, which compared the performance of

automatic systems to the manual diagnosis performed by the experts.

Figure 1.1 Methodology of the research

13

1.5 Thesis Overview The thesis is organised into six chapters, each focusing on different features of the

research work. The following is a summary of the contents of each chapter.

Chapter 1 provides an overview of diabetic retinopathy screening and a more detailed

investigation of the problems, particularly in Malaysia. The research aims and

objectives of this study are also presented. In addition, the motivations which have led

to this research, the contributions of the thesis and the research methodology are

presented in this introductory chapter.

Chapter 2 describes the background and the literature review in addition to basic

information on diabetes mellitus and diabetic retinopathy, as well as the prevalence of

diabetes mellitus and diabetic retinopathy worldwide and, particularly, in Malaysia. It

also provides information on the level of advancement in the area of image processing

for diabetic retinopathy screening systems. The chapter also highlights the

implementation of fuzzy image processing techniques on medical and non-medical

images, which is the core of this research work.

Chapter 3 discusses the research methodology used, including the process of data

collection for this study. The research design, as a guide for planning the research

development, is also presented. The experimental datasets, which consist of the existing

datasets and a novel developed dataset, are presented in Chapter 3 in greater detail. The

development of this new dataset is highlighted, including the expert diagnosis process,

the diagnosis summary and its overall advantages.

Chapter 4 explains the development of the proposed systems, using the previously

described datasets and image processing techniques. Each system presents a

combination of different techniques, such as different pre-processing techniques,

different feature parameters and different classifiers in the diabetic retinopathy

screening system. These systems are different from those proposed by other researchers.

The evaluations of the developed systems are also presented in Chapter 4, where it

14

presents the efficiency and the validity of diabetic retinopathy classification through the

developed systems.

Chapter 5 presents the overall results analysis of a new dataset. It also presents the

overall result analysis for the automatic developed systems. The chapter discusses the

analysis performed on the expert diagnosis, including descriptive and inferential

analysis. In order to generate a variance of system testing results and system

performance, the overall analysis of the developed systems are presented in two ways:

confusion matrix and statistical analysis. In addition, some discussions on the findings

of this study are presented.

Chapter 6 summarises the accomplishments of the research work. It concludes the

contents of the thesis and also highlights some recommendations for future research

work. It also provides information regarding the research contributions, which have

benefited a number of areas.

15

2 BACKGROUND AND LITERATURE REVIEW

Diabetes Mellitus (DM) is a significant public health concern. The diabetes epidemic is

leading to an increasing number of severe and chronic complications, including those

that are sight-threatening. Diabetic Retinopathy (DR) is a complication of diabetes

caused by high blood glucose. Diabetic retinopathy is a microvascular complication of

both insulin dependent (type 1) and non-insulin dependent (type 2) diabetes. It is one of

the diabetes mellitus complications that damages blood vessels inside the retina. The

retina is located at the back of the eye. Diabetic retinopathy commonly affects both eyes

and can lead to vision loss if it is not promptly treated (Centre for Eye Research

Australia, 2013).

This chapter provides the background for each of the main components involved in the

research. It comprises seven main components which are diabetes mellitus, diabetic

retinopathy, epidemiology of diabetic retinopathy, classification of diabetic retinopathy,

diabetic retinopathy screening, diabetic retinopathy image processing and finally, fuzzy

image processing, particularly on medical images. The chapter starts with the

explanation of diabetes mellitus in Section 2.1, followed by Section 2.2 which presents

one of its complications, i.e., diabetic retinopathy. Section 2.3 reveals the prevalence of

diabetes mellitus and diabetic retinopathy from a global perspective in addition to an

epidemiologic perspective in Malaysia. The classification of diabetic retinopathy is

explained in Section 2.4. Section 2.5 describes how diabetic retinopathy screening is

performed, including the grading process, examination schedule, diabetic retinopathy

treatment and management in addition to the follow up schedule. The image processing

approach, which is particularly used in diabetic retinopathy screening research, is

explained in Section 2.6. The implementation of fuzzy image processing that focuses on

medical images, which is the core of the proposed research, is presented in Section 2.7.

Finally, a summary of the second chapter is provided in Section 2.8.

2.1 Diabetes Mellitus Diabetes mellitus is a disorder caused by constant hyperglycemia of variable severity,

incidental to a lack or lessened efficacy of insulin (Ministry of Health Diabetic

16

Retinopathy Screening Team, 2012a). Meanwhile, Scanlon et al. (2012) have defined

diabetes mellitus as a chronic condition due to an excess of glucose circulating in the

bloodstream. Diabetes is a disorder caused by high levels of glucose in the blood

(Taylor and Batey, 2012; NHS Choices, 2012). It happens either when the pancreas

does not produce enough insulin or because cells do not respond to the insulin

produced. Insulin is a peptide hormone, produced by beta cells of the pancreas, a large

gland which is located behind the stomach. There are two types of diabetes mellitus,

which are Type 1 diabetes and Type 2 diabetes.

The increasing numbers of cases of diabetes are due to the following factors: a longer

life-span, modern lifestyles (urbanisation, mechanisation and industrialisation) and also

environmental and social factors, such as an unhealthy diet, obesity and physical

inactivity (International Diabetes Federation, World Health Organization and Secretariat

of the Pacific Community, 2000; Ministry of Health Diabetic Retinopathy Screening

Team, 2012a; Sivaprasad et al., 2012), in addition to uncontrolled hypertension and

smoking (Health Technology Assessment Unit, Medical Development Division,

Ministry of Health Malaysia, 2002). Tajunisah and others (2006) have claimed that the

duration of diabetes, hypertension and systemic complications including diabetic foot

ulcer, lower limb amputation, nephropathy and neuropathy were also factors of

retinopathy incidence. In addition, Mallika et al. (2011) confirmed that the duration of

diabetes, body mass index and visual loss are associated with diabetic retinopathy.

Meanwhile, the main symptoms of both types of diabetes are thirst, urinating frequently

(particularly at night), tiredness, weight loss, loss of muscle bulk, skin infections and

urinary infections (NHS Choices, 2012; Taylor and Batey, 2012).

Diabetes causes capillaries problems in the body, and the merely way to visualise this

condition is by looking into the retina (Taylor and Batey, 2012). Figure 2.1 shows the

anatomy of the eye including the retina, optic nerve, retinal vessels, cornea, lens, iris

and sclera. The retina is a light-sensitive tissue that located at the back of the eye, as

illustrated in Figure 2.1. Diabetes mellitus, if left untreated, can cause many health

problems. Among the systemic complications of diabetes mellitus are stroke,

17

cardiovascular disease, diabetic neuropathy, diabetic nephropathy and also diabetic

retinopathy (Ministry of Health Diabetic Retinopathy Screening Team, 2012a).

Figure 2.1 Retina (eyeSmart, 2014)

2.2 Diabetic Retinopathy The clinical manifestations of retinopathy are due to two basic pathophysiologic

mechanisms: increased capillary penetrability and the closure of the retinal capillaries

(Health Technology Assessment Unit, Medical Development Division Ministry of

Health Malaysia, 2002). Meanwhile, Taylor and Batey (2012) defined the term

‘retinopathy’ as a disease of the retina, and explained how diabetic retinopathy occurs.

High levels of blood sugar eventually cause capillary damage, where the lining cells of

capillaries become activated and ‘leaky’. Capillary closure or occlusion happens later,

due to the capillary damage and also the increase of the platelet stickiness and clotting

factors. As a result, the capillaries fail to supply nutrients to the retina as usual and

produce ischaemia, which is a decreased blood flow.

This item has been removed due to 3rd Party Copyright. The unabridged version of the thesis can be found in the Lancester

Library, Coventry University.

19

A fundus image with diabetic retinopathy features is presented in Figure 2.3, while the

explanation of the features is presented in Table 2.1. Amongst the detected retinopathy

signs are the microaneurysms, the retinal haemorrhages, the hard exudates, cotton wool

spots, abnormal new vessels and venous beadings, which are presented in Figure 2.3.

The definitions of these signs of diabetic retinopathy are listed in Table 2.1.

Table 2.1 Terminology definition of diabetic retinopathy signs (adapted from

Taylor and Batey, 2012)

This item has been removed due to 3rd Party Copyright. The unabridged version of the thesis can be found in the Lancester Library, Coventry University.

20

Table 2.1 (continued)

Figure 2.3 Features of diabetic retinopathy (Ministry of Health Diabetic

Retinopathy Screening Team, 2012b)


This


21

Figure 2.4 shows a comparison between the ordinary retina and a retina with diabetic

retinopathy signs. Figure 2.4 (a) shows the normal retina, which is free of any signs of

diabetic retinopathy. Meanwhile, Figure 2.4 (b) shows the retina with the presence of

several features of diabetic retinopathy, such as microaneurysms, haemorrhages,

exudates, cotton wool spots and maculopathy. The terminology of these features is

described in Table 2.1, while the visualisation of these features is presented in Figure

2.3.

(a) Normal retina

(b) Diabetic retinopathy signs

Figure 2.4 Normal retina compared to retina with diabetic retinopathy signs

2.3 Epidemiology of Diabetic Retinopathy

This section will reveal both the Malaysian and the worldwide prevalence of visual

impairment and diabetic retinopathy.

2.3.1 Diabetic Retinopathy and Global Epidemiology

Diabetic Retinopathy is an important cause which can lead to blindness. Global Data on

Visual Impairment, a 2010 article by the World Health Organization (WHO) stated that

globally, the number of visually impaired people of all ages was estimated to be 285

million, where 39 million are blind and 246 million suffered from low vision (World

Health Organization, 2012b). In addition, people aged 50 years and older make up 82%

This item has been removed due to 3rd Party Copyright. The

unabridged version of the thesis can be found in the Lancester Library, Coventry University.


unabridged version of the thesis can be found in the Lancester Library, Coventry University.

22

of the blind and 65% of the visually impaired. The report makes estimates for three age

groups (0 to 14 years, 15 to 49 years and 50 years above) coming from six WHO

regions (the African Region, the Region of the Americas, the Eastern Mediterranean

Region, the European Region, the South-East Asian Region and the Western Pacific

Region). The major global causes of visual impairment are uncorrected refractive errors

(42%), cataracts (33%), glaucoma (2%), age related macular degeneration (1%),

diabetic retinopathy (1%), trachoma (1%), corneal opacities (1%) and the remaining

18% is undetermined. Meanwhile, the major causes of blindness are cataracts (51%),

glaucoma (8%), age related macular degeneration (5%), childhood blindness (4%),

corneal opacities (4%), uncorrected refractive errors and trachoma (3%), diabetic

retinopathy (1%) and other undetermined causes (21%). The global visual impairment

and blindness reasons are depicted in pie chart form in Figure 2.5 and Figure 2.6,

respectively. The report summarises that retinal diseases are the main global of visual

impairment cause. In addition, the article confirms that the total number of visual

impairments and blindness due to age-related macular degeneration, glaucoma and

diabetic retinopathy is more compared to trachoma and corneal opacities, which are the

infective causes. The report suggests an urgent development of the eye care system

including rehabilitation, education and support services is required to overcome those

enduring eye diseases. It can be concluded therefore that diabetic retinopathy is among

of the visual impairment and blindness causes. Thus, this eye problem should be

addressed before it is worsens.

In addition, Sivaprasad and others (2012) examine a global prevalence of diabetic

retinopathy according to ethnicity and region. The survey reveals that the prevalence of

diabetic retinopathy, including sight threatening diabetic retinopathy and macular

edema, are higher in South Asian, African and Latin American populations compared to

the white population. The survey also concludes that ethnic-specificity is one of the

contribution rates of diabetic retinopathy, in addition to other factors, including the

length of exposure and severity of hyperglycemia, hypertension and hyperlipidemia. In

addition, factors like obesity, urbanisation, changes in diet, sedentary lifestyles and

communicable diseases rate will increase the demands on healthcare for many

ethnicities, particularly in Asia.

23

Figure 2.5 Global causes of visual impairment, including blindness, as

percentage (World Health Organization, 2012b:6)

Figure 2.6 Global causes of blindness as a percentage of global blindness

in 2010 (World Health Organization, 2012b:6)

2.3.2 Diabetic Retinopathy Epidemiology in Malaysia

Diabetes mellitus is a significant public health concern and one of the most established

enduring diseases in Malaysia. The National Health and Morbidity Survey (NHMS)

presented the diabetes prevalence for Malaysians (Letchuman et al., 2010). The survey,

This item has been removed due to 3rd Party Copyright. The unabridged version of the thesis can be found in the Lancester Library, Coventry

University.


University.

24

which involved 34,539 respondents and covered all states of Malaysia, reported that the

highest prevalence of diabetes among ethnic groups is thus: Indians (19.9%), followed

by Malays (11.9%) and Chinese (11.4%). In addition, the survey also revealed other

information, such as the prevalence of people with known diabetes was 7.0%, while

newly diagnosed diabetes was 4.5% and 4.2% for Impaired Fasting Glycaemia. It was

also reported that about 73.5% of the patients used the government healthcare services

and only 45.05% of known diabetics had undergone an eye examination. Letchuman

and others (2010) were unsatisfied with the results showing that only one third of

respondents had their eyes examined within one year from the time of the survey.

Although they are required to have an eye examination at diagnosis and subsequently

annually (as recommended by the clinical practice guidelines on the management of

diabetic retinopathy in type 2 diabetes), the patients seldom had their eye examinations

as suggested. In addition, Letchuman and others (2010) conclude that the reasons for

the lower percentage of eye examination are include: patients’ lack of awareness of the

schedule and the lack of eye screening services due to insufficient numbers of trained

staff and resources.

The survey also reported the prevalence of diabetes cases in Malaysia by state, where

Negeri Sembilan, Melaka and Penang have the highest frequency of diabetes, with

15.3%, 15.2% and 14.9%, respectively. The location of this research is Melaka, which

is highlighted as having the highest prevalence of diabetes. The survey also reveals that

the prevalence of diabetes is higher in urban (12.2%) compared to rural areas (10.6%).

In addition, Mafauzy and colleagues (2011) proposed a study on the diabetes care status

in Malaysia, which involved 1,670 patients from general hospitals, diabetes clinics and

referral clinics. The most commonly reported eye complications of the study include

proliferative diabetic retinopathy.

Goh’s research (2008) reported that in 2007 the total number of diabetics with diabetic

retinopathy registered to the Diabetic Eye Registry was about 36.8%. In addition, the

research revealed some demographic characteristics, such as the mean age of the

registered patients was 57.2 years and that half of them (52.8%) were aged between 30

to 60 years. Diabetic patients were most commonly female at 54.6% and the majority of

25

the patients were Malay (54.0%). Moreover, about 92.2% of patients diagnosed with

type 2 diabetes mellitus had the condition for below 10 years and 91.7% had been

referred from a government healthcare facility. It was also reported that two thirds of

diabetic patients had never previously undergone an eye examination and 71.9% of

those who had an eye examination had attended a year ago. There were about 63.3%

patients diagnosed as having no diabetic retinopathy, in addition to mild non-

proliferative diabetic retinopathy accounting for 16.5%, moderate non-proliferative

diabetic retinopathy (9.8%), severe non-proliferative diabetic retinopathy (3.4%),

proliferative diabetic retinopathy (7.1%) and maculopathy was detected in 9.5% of

cases (Goh, 2008).

The 7th

Report of the National Eye Database 2013 by the NED Steering Committee

Members (2015) presented the distribution of three eye diseases, which consisted of

diabetic retinopathy, glaucoma and lens-related complications, from the year 2002 to

2013, as illustrated in Figure 2.7. It can be concluded that among the three eye diseases,

diabetic retinopathy showed the highest percentage of distribution. The percentage of

diabetic retinopathy cases however, seemed to be reducing from 2010 onwards.

Figure 2.7 Percentage distribution of diabetic retinopathy cases, glaucoma or

lens-induced glaucoma, 2002-2013 (NED Steering Committee Members,

2015:112)


26

In addition, Addoor and others (2011) conducted a study of the awareness of diabetic

retinopathy among the patients attending the diabetic clinics in Melaka, Malaysia and

reported that 79.8% of the respondents were aware of diabetes mellitus complications

and 87.2% were aware that diabetes can affect the eyes. The study also reported

however that only 50% of the patients had undergone an ophthalmological evaluation.

The proposed study concluded that although awareness among the patients was good,

the motivation to undergo the assessment was poor. Furthermore, Thevi et al. (2012)

revealed the eye diseases and visual impairment frequency among the rural population.

It was reported that cataracts (22.9%) were the most popular eye disease, after that

retinal illnesses including diabetic retinopathy (11.5%) and ocular trauma (9.8%).

2.4 Classification of Diabetic Retinopathy There are several classifications or stages of diabetic retinopathy. The Screening for

Diabetic Retinopathy Report by the Ministry of Health in Malaysia explained the

technical features of how the stages of diabetic retinopathy occur (Health Technology

Assessment Unit, Medical Development Division, Ministry of Health Malaysia, 2002).

The enlargement of the veins in the retina is reported as one of the initial signs of

diabetic retinopathy. The small capillaries present may also produce some changes,

which affect the blocking, and later result in small swellings in the vascular walls,

called microaneurysms. This is termed minimal non-proliferative diabetic retinopathy

(NPDR), and, at this early stage, the sight may not be affected. The incidence of

diabetic retinopathy continued in cases where the blood flow gradually worsens and

progressively causes damage to larger portions of the retina. At this level, small

haemorrhages and other vascular changes in the fundus of the eye appear, due to

vascular blocking and leakage. Consequently, the retinopathy progresses from minimal

to mild, due to the appearance of retinal haemorrhages, hard exudates and nerve layer

infarct. In addition to minimal and mild NPDR, moderate NPDR occurs due to venous

beading and intra-retinal microvascular irregularities. For severe NPDR classification,

there should be more haemorrhages, microaneurysms, intra-retinal microvascular

abnormalities and venous beading than is present in moderate NPDR (Health

Technology Assessment Unit, Medical Development Division, Ministry of Health

Malaysia, 2002).

27

The next severe stage of diabetic retinopathy, or proliferate diabetic retinopathy (PDR)

occurs when new vessels (neovascularisation) are detected. This stage is high risk and

could result in blindness. These diabetic retinopathy stages and their respective features

are presented in Table 2.4. The macula disorders which affect the central vision happen

due to visual loss from diabetic retinopathy. Therefore, eye examinations for the

diagnosis of diabetic retinopathy at an initial stage are important, because diabetic

retinopathy can bring to blindness. Moreover, regular screening can help save sight.

2.5 Diabetic Retinopathy Screening

Screening is defined as the testing of a population in order to identify individuals

exhibiting attributes that could be early symptoms, or indicators, of a predisposition

associated with a particular condition (Taylor and Batey, 2012). The main purpose of

diabetic retinopathy screening is to detect whether the individuals require referral for

further treatment, in order to prevent blindness (Taylor and Batey, 2012). In addition to

this main purpose, there are other purposes for diabetic retinopathy screening. These

include: identifying the disease at an early stage; possibly detecting a requirement for

blood pressure and blood sugar treatment; educating the population about the causes of

diabetic retinopathy and ways to reduce the risk; and, potentially, to identify non-

diabetic conditions through the screening process (Taylor and Batey, 2012). One major

problem is that diabetic eye disease does not interfere with sight until it reaches an

advanced stage (Taylor and Batey, 2012). Laser treatment can save sight, but only if it is

used at an early stage. This shows the importance of essential regular screening, which

can help detect the diabetic patients at an early stage of diabetic retinopathy. Moreover,

earlier identification of any retinopathy signs can allow changes to blood pressure or

blood glucose management in order to slow the rate of progression.

Digital retinal imaging is usually used as a screening technique. The proposed method is

recommended in the UK and also by the National Retinal Screening Project Group

(Taylor and Batey, 2012). However, there are some challenges, such as efficient and

cost-effective measures, that need to be considered to establish the screening systems.

Furthermore, Taylor and Batey (2012) considered digital retinal imaging as one of the

screening choices and underlined five principles of retinal screening. These are

28

comprised of regular screening assurance, the availability of an efficient screening

system, eye screening practise being included as part of diabetes care, the

ophthalmologist’s participation in the planning and operation of the screening system

and finally, the quality control of the screening process. Scanlon and colleagues (2009)

recommended four steps in systematic screening programmes developed for the sight-

threatening diabetic retinopathy. Firstly effective treatment, opportunistic as well as

systematic screening, and finally, full quality assurance and coverage screening.

Meanwhile, Hutchins et al. (2012) presented a study of diabetic retinopathy screening in

New Zealand, and concluded that among the requirements in improving the retinal

screening quality should be quality data and quality assurance platforms.

As discussed in Chapter 1, there are many screening methods available for diabetic

retinopathy screening. Figure 2.8 shows the screening modalities used in screening

programmes including the direct ophthalmoscope, the slit-lamp biomicroscopy with

contact lens, the binocular indirect ophthalmoscopy and the fundus photography.

However, different screening modalities will provide a variation in the sensitivities and

specificities obtained.

Table 2.2 describes the diagnostic accuracy of different screening tools, which are the

direct ophthalmoscope, the slit lamp biomicroscope, the mydriatic fundus camera and

also the non-mydriatic fundus camera. It can be concluded from Table 2.2, that the non-

mydriatic fundus camera has high sensitivity and specificity, eliminating the need for

pupillary dilatation, promoting compliance, efficiency and safety. The clinical practice

guidelines for the screening of diabetic retinopathy (Ministry of Health Malaysia,

Malaysian Society of Ophthalmology and Academy of Medicine Malaysia, 2011)

recommend that the non-mydriatic fundus camera should be used as a screening tool for

diabetic retinopathy, whenever possible with a double field fundus photo assessment.

When there is no access to a fundus camera, an ophthalmoscope should be used.

29

Direct ophthalmoscope Slit-lamp biomicroscopy with contact

lens

Binocular indirect ophthalmoscopy Fundus photography

Figure 2.8 Diabetic retinopathy screening tools (Ministry of Health Diabetic

Retinopathy Screening Team, 2012c:14)

Table 2.2 Sensitivity and specificity of diabetic retinopathy tools (Ministry of

Health Malaysia, Malaysia Society of Ophthalmology and Academy of Medicine

of Malaysia, 2011)

This item has been removed due to 3rd Party Copyright. The unabridged version of

the thesis can be found in the Lancester Library, Coventry University.








30

Generally, in current practice, all registered diabetic patients will first undergo the

visual acuity test. Following this, fundus photography session will be organised using

the fundus camera and assessment will then be performed. A trained paramedic will

play a role in grading the captured fundus images into a diabetic retinopathy stage,

based on the changes or abnormalities shown on the fundus images. The management of

the condition, including a follow-up schedule or referral to an ophthalmologist for

further diagnosis, will be decided upon based on the retinopathy stages. This

explanation of the diabetic retinopathy screening process is illustrated in Figure 2.9.

Currently in Malaysia, the clinicians use the non-mydratic fundus camera to capture

retinal images. The fundus camera comprises an internal or external digital camera, a

computer and a software system. It is used because of its strengths as it is easy to use,

patient-friendly, user-friendly, time and cost effective and also pupil dilatation is

performed only if necessary. The fundus camera produces high quality digital

photographs that can be viewed immediately and shown to the patients to improve their

understanding of the disease. A fundus camera is designed to photograph the interior

surface of the eye such as the retina, the optic disc, the macula, and the posterior pole

(i.e., the fundus). Fundus cameras are used by optometrists, ophthalmologists, and

trained medical professionals in order to monitor the disease progression, for the

diagnosis of a disease or in screening programmes, where the photos can be analysed

later.

31

Figure 2.9 Screening process of diabetic retinopathy to prevent blindness

(Ministry of Health Malaysia, Malaysian Society of Ophthalmology and Academy

of Medicine of Malaysia, 2011)

The handbook or guide to diabetic retinopathy screening by the Ministry of Health

Diabetic Retinopathy Screening Team (2012b) explains the technique of photography

and also details the features of diabetic retinopathy, visible through the use of the

fundus camera. The screening group suggested that two photographs of two views (with

the optic disc as the centre and the fovea as the centre) should be taken as the input for


32

screening, as presented in Figure 2.10. The optic disc or the optic nerve is a round area

in the back of the eye while the fovea is the centre of the macula region. Figure 2.11

presents good quality fundus photos, which are essential for the process of diabetic

retinopathy screening.

Optic disc as the centre Fovea as the centre

Figure 2.10 Fundus camera photography (Ministry of Health Diabetic Retinopathy

Screening Team, 2012b, 2012d)

Figure 2.11 Good quality fundus photos (Ministry of Health Diabetic Retinopathy

Screening Team, 2012b)

This

This


unabridged version of the thesis can be found in the

Lancester Library, Coventry University.











33

The fundus image can be divided into four quadrants which are superonasal,

inferonasal, superotemporal and inferotemporal, as shown in Figure 2.12. The macula

region is at the posterior pole of the eye, within one disc diameter of the fovea (Taylor

and Batey, 2012).

Fundus image quadrants Macula

Figure 2.12 Normal fundus image (Ministry of Health Diabetic Retinopathy

Screening Team, 2012b)

2.5.1 Diabetic Retinopathy Grading

Based on the fundus image captured by the fundus camera, the experienced paramedic

will then make the diagnosis as to whether or not the patient has any eye disease,

including diabetic retinopathy. The grading is performed manually by the paramedics,

with reference to any changes or abnormality seen on the retinal image. Table 2.3 shows

the findings of diabetic retinopathy features for each stage of the condition. The

presented retinopathy stages are the non-proliferative diabetic retinopathy (mild,

moderate and severe), the proliferative diabetic retinopathy, the diabetic maculopathy

(mild, moderate and severe) and the advanced diabetic eye disease. Information about

the retinopathy stages and the terminology used to describe their features of are

presented in Section 2.4 and Table 2.1, respectively, for a detailed explanation.

This item has been removed due to 3rd Party Copyright. The unabridged version of the thesis can be found in

the Lancester Library, Coventry University.



37

The Early Treatment Diabetic Retinopathy Study (ETDRS) group initially proposed the

diabetic retinopathy severity scale (Early Treatment Diabetic Retinopathy Study

Research Group, 1991). In addition to the early classification by ETDRS, an

international clinical disease severity scale for diabetic retinopathy and diabetic macula

oedema (DME) was proposed by Wilkinson and others in 2003. It is much simpler,

having less severity levels and diagnostic criteria as compared to the ETDRS

classification system. Table 2.4 shows the Diabetic Retinopathy Disease Severity Scale

including no apparent retinopathy, nonproliferative diabetic retinopathy (mild, moderate

and severe) and proliferative diabetic retinopathy. Table 2.5 shows the Diabetic Macular

Edema Disease Severity Scale proposed by Wilkinson and colleagues (2003), which has

been used as an international scale in order to assist in the grading of fundus images into

distinct categories based on the retinal findings.

Table 2.4 International Clinical Diabetic Retinopathy Disease Severity Scale

(Wilkinson et al., 2003)


38

Table 2.5 Diabetic Macular Edema Disease Severity Scale (Wilkinson et al., 2003)

2.5.2 Examination Schedule

Regular screening is important for early detection and early treatment, in order to assist

the management of diabetic retinopathy. The preliminary fundus examination for

diabetic retinopathy varies according to the types of diabetes mellitus, and is presented

in Table 2.6 below.

Individuals with diabetes mellitus should be screened at least every two years.

Individuals who can be categorised as high risk (i.e. with a longer duration of diabetes

or poor control of blood sugar, blood pressure or serum lipid) should be examined at

least annually. Moreover, individuals with any signs of non-proliferative diabetic

retinopathy should be examined at 6-12 monthly intervals. Earlier follow-up may be

required in high risk groups, including the presence of renal complications and the

progression of diabetic retinopathy. Table 2.7 provides the recommendations for eye

examination follow-up schedules for patients with diabetes mellitus.


39

Table 2.6 Timing of the first screening (Ministry of Health Malaysia, Malaysian

Society of Ophthalmology and Academy of Medicine of Malaysia, 2011)

Table 2.7 Recommended follow-up schedule (American Academy of

Ophthalmology Retina Panel, 2008)

The crucial aim for diabetic retinopathy screening is to detect sight threatening factors

and to ensure their timely treatment in order to prevent vision loss. Therefore,

appropriate referral to an ophthalmologist should be performed. Among the criteria

which would require referral are the presence of any level of diabetic maculopathy,

severe NPDR, any PDR, unexplained loss of vision and also if the screening

examination cannot be performed with an upgradable fundus photo. The National

Institute for Clinical Excellence (NICE) presents the urgency of referral as shown in

Table 2.8. It is recommended that the examination schedule and urgency of referral to

an ophthalmologist should be based on the grade and severity of the diabetic retinopathy

along with the presence of risk factors (Ministry of Health Malaysia, Malaysian Society

of Ophthalmology and Academy of Medicine of Malaysia, 2011).



40

Table 2.8 Criteria for urgent referral (National Institute for Clinical Excellence,

2002)

2.5.3 Diabetic Retinopathy Management and Treatment

Early detection of diabetic retinopathy is vital because diabetic retinopathy is

irreversible. The standard practice for treating diabetic retinopathy is by using laser

photocoagulation, where the blood capillaries immersed the energy. The summary of

the treatment for diabetic retinopathy is presented in Table 2.9.

Table 2.9 Summary of treatment for diabetic retinopathy (Ministry of Health

Malaysia, Malaysian Society of Ophthalmology and Academy of Medicine of

Malaysia, 2011)



41

2.6 Diabetic Retinopathy Image Processing

Currently, image processing techniques are widely used as a means of diagnosing

diseases, including eye diseases. Computer-based imaging tools are necessary to

effectively detect signs of diabetic retinopathy. Early detection would allow the

ophthalmologist to treat patients before major damage occurs and would present the

best chance of protecting the patient’s vision. The automatic diabetic retinopathy

grading system would allow a faster and more efficient diagnosis. Preventive actions

could be taken early to protect vision and avoid blindness.

Diabetic retinopathy screening is currently a common research area, in which some

researchers focus on finding and proposing several techniques or methods for detecting

certain features of diabetic retinopathy (i.e., microaneurysms, haemorrhages, exudates,

and neovascularisation). Nonetheless, there are some researchers who propose the

development of automated systems for detecting and classifying normal or abnormal

diabetic retinopathy.

Digital image processing systems generally have three main parts: image preprocessing,

feature extraction and classification.

2.6.1 Image Preprocessing

Preprocessing is the process of enhancing or improving features of image data for the

next processing task. According to Sonka and colleagues (2008), image preprocessing

methods can be classified into four categories: pixel brightness transformations,

geometric transformations, local preprocessing and image restoration. However, some

papers classify image preprocessing methods into image enhancement and image

restoration only.

Pixel brightness transformations deal with pixel brightness, and consist of brightness

corrections and greyscale transformations. The brightness of pixel is adjusted in

brightness correction according to the current brightness and image location; whereas in

greyscale transformation, the brightness values contrast of the image is enhanced. The

greyscale image is defined as a data matrix which the values are indicated by the shades

42

of grey (Gonzalez et al., 2009). Sonka and others (2008) concluded that greyscale

transformations are mainly used for manual viewing in which the image is simply

defined in an improved contrast. For example, greyscale transformation provides a clear

contrast to an X-ray image. Greyscale transformation technique includes histogram

equalisation for enhancement of contrast. The technique creates an equal distribution of

brightness level for the image.

Geometric transformations offer the removal of geometric misrepresentation that

happens during the image capturing process. The two basic steps of a geometric

transform are pixel coordinate transformation and brightness interpolation. In addition,

a significant purpose of image restoration methods, are to subdue degradation. The

image restoration methods apply the concept of deconvolution across the whole image.

Local preprocessing methods utilise a small pixel neighbourhood in order to generate an

output image with a new brightness value. Two common groups are used to achieve

this, namely smoothing and edge detection. Smoothing is used to reduce noise or other

minor fluctuations in the image. Gradient operators determine edges where the locations

undergo fast changes. There are two components of edge detection, which are

magnitude and direction. Most gradient operators such as Roberts, Laplace, Prewitt,

Sobel, Robinson and Kirsch can be expressed using convolution masks.

Amongst the preprocessing techniques used in the present diabetic retinopathy detection

system are greyscale conversion, green channel extraction, contrast enhancement (such

as histogram equalisation), filtering, morphological operations, segmentation and

thresholding among others. Within the scope of diabetic retinopathy detection, optic

disc elimination and blood vessel removal are two main processes that are widely used

as an additional stage, before the next process is performed. In addition, the localisation

of the fovea and macula are important for the detection of maculopathy. This detection

of these retinal structures also requires some preprocessing techniques. Table 2.10 lists

the preprocessing techniques which have been implemented in diabetic retinopathy

detection research.

52

2.6.3.1 Support Vector Machines

Support Vector Machines (SVM) was introduced in 1992 by Boser, Guyon and

Vapnikin in COLT-92 (Jakkula, 2008). SVM are a set of related supervised learning

methods used for classification and regression. A support vector machine is a

classification and regression prediction tool that uses machine learning theory to

maximise the predictive accuracy while automatically avoiding the over-fit of the data

(Jakkula, 2008). SVMs can be defined as systems which use the hypothesis space of a

linear function in a high dimensional feature space. The original support vector

classifier was developed for the linear separation of two classes and later this limitation

was overcome by allowing non-linearly separable classes, non-separable classes,

combining multiple 2-class classifiers to yield multi-class classification and other

extensions (Sonka et al., 2008).

An SVM works by finding the best hyperplane which separates all data points, which is

where the best hyperplane will have the largest margin between the two classes. The

margin is the maximal width of the slab parallels to the hyperplane that has no interior

data points. The data points closest to the separating hyperplane are support vectors.

Figure 2.13 illustrates the functioning of SVM.

An SVM is a popular technique for data classification. Jakkula (2008) concluded its

strengths and weaknesses. The major strength of an SVM is that the training is

relatively easy, with no local optimum. In addition, it scales relatively easy to high

dimensional data and the trade-off between classifier complexity and error can be

controlled explicitly. On the other hand, SVM requires a good kernel function and this

is one of its weaknesses (Jakkula, 2008; Burges, 1998). In addition its speed and size

are its other limitations, both in training and testing (Burges, 1998).

The SVM has been successfully used for pattern classification problems and is also

widely used in medical image processing and analysis. Mookiah and others (2012)

proposed data mining techniques for an automated diagnosis of glaucoma using Higher

Order Spectra (HOS) and Discrete Wavelet Transform (DWT) features. In this case, a

support vector machine classifier is used to identify the glaucoma and normal images

automatically and with high accuracy. Priya and Aruna (2011) investigated and

53

proposed a computer-based system for diabetic retinopathy in identifying normal,

NPDR and PDR classes. The proposed system uses colour fundus images, where the

features are extracted from the raw image using image processing techniques and fed to

a support vector machine for classification. Experimental results show that the

classification accuracy can provide an improved result, with a sensitivity of 99.45% and

a specificity of 100%.

Figure 2.13 Basic two-class classifications with Support Vector Machines (The

Mathworks, Inc., 2016a)

2.6.3.2 Neural Networks

Neural networks are widely used in application areas such as system identification and

control, pattern recognition, medical diagnosis, financial application, data mining,

visualisation and many more. One of the main applications of neural networks is

classification. A neural network consists of many neurons called units or nodes, each of

which performs two functions. Firstly the aggregation of its input from other neurons or

the external environment, and secondly, the generation of an output from the aggregated

inputs (Er et al., 2003).



54

According to Er and others (2003), a neural network is characterised by its architecture,

or connecting pattern among the neurons, its method of determining the weights of the

connections and its activation function. Er and colleagues (2003) also presented

biomedical engineering applications using one of the Artificial Neural Network (ANN)

paradigms. One of its popular applications is in the classification of breast cancer. Er

and others (2003) presented the methods and results of the studies corresponding to

extracting rules from a mammography dataset and automatically constructed a fuzzy

classifier using the Dynamic Fuzzy Neural Network (D-FNN). This was done in order

to classify the two important features in breast cancer diagnosis, which are benign and

malignant masses. Lin and other co-workers (2003) proposed a neural networks method

to determine the progression of glaucoma based on visual field thresholds and

concluded that the progression of glaucoma could be detected from visual field

thresholds with a neural network.

Gardner and others (1996) initially developed a screening tool for the automatic

detection of diabetic retinopathy using an artificial neural network and comparing the

network against the ophthalmologist screening of a set of fundus images. As a result,

the network achieved good accuracy therefore the system could be used as an aid to the

screening of diabetic retinopathy for diabetic patients. Priya and Aruna (2012) enhanced

the computer-based system for diabetic retinopathy in identifying normal, NPDR and

PDR classes, by using two types of classifiers: a Probabilistic Neural Network (PNN)

and a Support Vector Machine (SVM). The classifiers are described in detail and their

performances are compared. As a conclusion, it is shown that, from the results obtained,

the SVM model is more effective compared to the PNN. Priya and Aruna (2013a)

proposed and compared three models, namely a Bayesian classifier, while maintaining

the PNN and the SVM in the developed system. Experimental results show that the

SVM outperforms all other models and this proves once again, that the SVM is a better

choice in detecting and classifying DR categories. The detection of DR disease and its

classification with the help of the Radial Basis Function Neural Network (RBFNN)

method has also been proposed in (Priya et al., 2013b). However, the experimental

results show that the accuracy of the proposed system is relatively low (76.25%) and it

55

is recommended that its accuracy could be improved by finding more relevant features

and by combining it with other classification methods.

2.6.3.3 Other Classification Methods

Naive Bayes classification is another machine learning technique used for classification

purpose. The Naïve Bayes classifier can be used when the features are independent of

one another within each class, but it works well in practice even when that

independence assumption does not hold (The MathWorks, Inc., 2016b). The classifier

works in two steps. In the training step, the method estimates the parameters of a

probability distribution from the training samples by assuming that the features are

conditionally independent, given the class. The method computes the posterior

probability of an unseen test sample belonging to each class in the prediction step. The

method later classifies the test sample according to the largest posterior probability.

Decision trees are considered to be one of the most popular approaches for

implementing classifiers. A decision tree is a classifier in the form of a tree structure

and it classifies instances or examples by starting at the root of the tree and moving

through it until a leaf node is reached. Rokash and Maimon (2005) pointed out some

advantages of using a decision tree, for example it is easy to follow when compacted, as

it is self-explanatory. It can handle both nominal and numeric input attributes, can

represent any discrete-value classifier and is also capable of handling datasets with

errors and missing values.

In the k-nearest neighbour (k-NN) classifier, the object is classified by a ‘majority vote’

from its neighbours, with the object being assigned to the most common class among its

k nearest neighbours. K-NN offers several advantages including simplicity,

effectiveness, intuitiveness and competitive classification performance and it can handle

the noisy training data. However, the limitations of the k-NN method include a poor

run-time performance if the training set is large and it is also very sensitive to irrelevant

or redundant features because all the features contribute to the similarity and the

classification (Imandoust and Bolandraftar, 2013).

59

Dynamic Histogram Equalization (BPDHE) proposed by Ibrahim and Kong (2007), in

which the Gaussian kernel was used for a global image histogram smoothing, followed

by the segmentation of the valley regions for dynamic equalisation. However, this

technique processes the crisp statistics of digital images to enhance contrast, which

suffers from the limitation that it does not take into account the inexactness of grey-

values and also that crisp histograms need smoothing for equalisation partitioning. In

order to overcome this limitation, the fuzzy histogram is introduced to handle the

imprecision in grey levels with the appropriate fuzzy membership function, resulting in

there being no missing intensity levels and no random fluctuations. The use of fuzzy

statistics has improved the algorithm’s performance and improved its ability to preserve

brightness and provide better contrast enhancement as compared to BPDHE. The

proposed technique shows that it can preserve image brightness better than the

histogram equalisation and the other techniques based on contrast limited adaptive

histogram equalisation.

Figure 2.14 Brightness Preserving Dynamic Fuzzy Histogram Equalization stages

Fuzzy Histogram Computation

Partitioning of the Histogram

Dynamic Equalisation of the Histogram Partitions

Normalisation of Image Brightness

Contrast Enhanced Image

Low Contrast Image

60

Li et al. (2011) proposed a novel fuzzy level set algorithm for medical image

segmentation based on the segmentation obtained by spatial fuzzy clustering. The

proposed algorithm leads to a better segmentation and effectiveness for medical image

segmentation tasks. In addition, fuzzy filter techniques, which aim to detect and remove

the noise from the corrupted image, are proposed in (Toh et al., 2010; Kwan et al.,

2002; Kwan, 2003; Toh et al., 2008). Toh et al. (2008) proposed a new fuzzy switching

median (FSM) filter technique in image processing, which is an extension to the classic

switching median filter through employing a fuzzy mechanism in detecting the noisy

pixel. The proposed filter was able to effectively remove noise while preserving image

details and textures. Later, Toh and others (2010) proposed a novel two-stage noise

adaptive fuzzy switching median (NAFSM) filter for the detection and removal of salt-

and-pepper noise. The first stage, which is the detection stage, uses the histogram of the

corrupted image to identify noise pixels. Meanwhile, the second stage is the noise pixels

filtering. After that, fuzzy reasoning is applied to NAFSM filtering in order to handle

uncertainty present in the extracted local information. The proposed filter is able to

reduce high-density salt-and-pepper noise, as well as preserving fine image details,

edges and textures. Kwan (2002, 2003) proposed seven fuzzy filters for noise reduction

in images, where the fuzzy filters apply a weighted membership function to an image in

order to determine the centre pixel.

Another fuzzy image processing that can be implemented is fuzzy edge detection. An

edge is a boundary between two uniform regions. Therefore, the membership functions

are defined for fuzzy edge detection to state the degree of a pixel whether it represents

an edge or a uniform region. Figure 2.15 shows the membership functions example of

the inputs (image gradients) and outputs (intensity of the edge-detected image) for the

edge detection, while Figure 2.16 shows the output after the edge detection using fuzzy

logic.

61

Figure 2.15 Membership functions of the inputs and outputs (The Mathworks,

Inc., 2016c)

Figure 2.16 Edge detection using fuzzy logic output (The Mathworks, Inc., 2016c)

Original Greyscale Image Edge Detection Image


University.





62

2.8 Summary This chapter reviewed the prevalence of diabetes and diabetic retinopathy, including the

diabetic retinopathy incidence in addition to the signs and features of diabetic

retinopathy. The epidemiology of diabetic retinopathy is also presented to reveal the

severity of diabetic retinopathy disease. This chapter also summarised the classification

of and screening for diabetic retinopathy, in addition to diabetic retinopathy treatment

and the follow up schedule. Image processing was discussed as a tool to diagnose

diabetic retinopathy in a more precise and efficient way. At the close of the chapter,

background information about digital image processing, image analysis and processing

techniques (particularly in diabetic retinopathy detection research) were presented.

Finally, fuzzy image processing techniques were also presented.

The summary of preprocessing techniques, feature extractions methods and machine

learning techniques for classification used in previous diabetic retinopathy screening

research are presented in Table 2.10, Table 2.11 and Table 2.12, respectively.

Meanwhile, the fuzzy image processing techniques are presented in Section 2.7. It can

be concluded that various image preprocessing techniques, feature extractions and

machine learning techniques were proposed in order to produce an efficient and reliable

diabetic retinopathy system. The fuzzy image processing techniques help in generating

better quality of image and enhanced performance. However, fuzzy processing has not

been used during the preprocessing stage for diabetic retinopathy screening system

which involves fundus images. Therefore, this proposed research implements fuzzy

techniques for the image pre-processing part within the microaneurysms and

maculopathy detection.

63

3 RESEARCH METHODOLOGY AND DATA

COLLECTION

This chapter introduces the system design of the proposed research in Section 3.1.

Section 3.2 presents the target population involved in the research and the location of

the research study in Section 3.3. Section 3.4 explains the types of data collected and

the data collection procedure is presented in detail in Section 3.5. An analysis of the

data collected is presented in Section 3.6. Alongside the data collection and analysis

phase, ethical issues should be considered, which are outlined in Section 3.7. Finally, in

addition to the research data set collected, the benchmark public data sets used in the

proposed research are presented in Section 3.8 and followed by the summary of the

research methodology and data collection chapter in Section 3.9.

3.1 Research Design

Generally, there are six steps in the process of research proposed by Creswell (2012)

and they entail: recognising the research problem, review of related literature,

specifying a purpose for research, data collection, and analysis and reporting of data,

and finally, evaluating the research. Awang (2012) presented a flowchart of research

which although similar to Creswell’s is more general in its nature.

In general, the data collection techniques for this research project are as follows:

i. Personal interview: face to face informal interviews with the

ophthalmologists and trained healthcare staff involve in diabetic retinopathy

screening.

ii. Observation: directly observing the process of fundus images capture and its

manual grading by trained health care staff.

iii. Internal sources: extracting the existing patient’s folder from the personal

computer attached to the fundus camera which contains the patient’s details,

fundus images and diagnosis.

Figure 3.1 presents the research design for the development of this research on

‘Automatic Screening and Classification of Diabetic Retinopathy Eye Fundus Images’.

65

3.2 Study Population The study includes both type 1 diabetes mellitus (T1DM) and type 2 diabetes mellitus

(T2DM) patients. It also covers all individuals with diabetes mellitus including adults,

children and adolescents. The diabetes mellitus duration is associated to the prevalence

of diabetic retinopathy and it varies among nations and ethnicity. Therefore, this study

involves individuals from all ethnic groups in Malaysia.

3.3 Location of the Research Study The study area of Melaka state, is one of the fourteen states that make up Malaysia. The

state of Melaka is located in Peninsular Malaysia and is surrounded by Negeri

Sembilan, Johor and also the Straits of Malacca. The data is collected from the main

hospital in the state of Melaka, the Melaka Hospital, a government-funded public

hospital located at Jalan Mufti Haji Khalil, Melaka, Malaysia. It was chosen because its

demographic is diverse in terms of ethnicity-religion and socio-economic groupings in

the southern part of Malaysia. Generally, the Melaka Hospital is a referral hub for

patients from Melaka health centres in addition to other health centres in the north of

Johor and a district in Negeri Sembilan, which is Tampin. Furthermore, very few

studies of the prevalence of diabetes mellitus and diabetic retinopathy have been carried

out in the state. These have mainly focussed on larger urban areas such as Kuala

Lumpur and Putrajaya. Therefore, the study will indirectly reveal the diabetes mellitus

and diabetic retinopathy prevalence throughout the Melaka state, in order to bridge the

prevalence gap with respect to the diabetes mellitus and diabetic retinopathy

epidemiology in Malaysia. Figure 3.2 shows the site of the research: at the Melaka

Hospital, Malaysia.

3.4 Data Collection

This section introduces the types of data collected in this research project, which were

used to response to the research questions and to achieve the objectives. After defining

and identifying the research problem and the research design, the task of data collection

took place. The method of data collection to be used for the study deals with two types

of data: primary and secondary.

66

Main entrance of Melaka Hospital Eye Clinic, Department of Ophthalmology,

Melaka Hospital

Figure 3.2 Research study location

3.4.1 Primary Data

Data which is new, original and collected for the first time is termed as primary data

(Kothari, 2004). The Social Dimensions of the Watershed Planning (2006) defined

primary data as generated and compiled data through interviews, surveys or focus

groups. In addition, the Social Dimensions of the Watershed Planning (2006) also

claimed that those data types are designed in order to address an unavailable issue or

information need in any present sources. Primary data is tailored to provide well-

focused and exclusive support for the decision-makers of organisations, but on the other

hand, the collection and analysing of the data more expensive and time consuming

(Malhotra and Birks, 2006). In order to develop the proposed system for the purpose of

this project, primary data which is information gathered from observation and interview,

has been collected. The detail of the observation strategy is explained in Section 3.5.1,

while the interview strategy is explained in Section 3.5.2.

3.4.2 Secondary Data

Secondary data is defined as data which has been collected by another party and has

undergone a statistical process (Kothari, 2004). Secondary data is also classed as data

67

that has already been collected for some other purpose (Sounders et al., 2009; Malhotra

et al., 2006).

According to Sounders et al. (2009), organisations collect and store a diversity of data

in order to support their operations. Sounders et al. (2009) highlighted some of the

advantages of using secondary data in a research project, and they include having fewer

resource requirements, being inconspicuous, being practical for longitudinal studies and

offering comparative and contextual data. Sounders et al. (2009) however, also revealed

the disadvantages of using secondary data: data collected may not match with need,

access can be difficult or costly, in addition to unsuitable aggregations and definitions.

The secondary data for this research project however are collected from an internal

source, namely fundus images. Section 3.5.3 will describe the details of the secondary

data collection method.

3.5 Collecting Data Procedure

The data collection techniques for the primary data are observations and interviews,

while the secondary data is collected from existing fundus images.

3.5.1 Collecting Primary Data through Observation

Observation is one of the techniques implemented for this study. According to Sounders

et al. (2009), participant observation and structured observation are two types of

observation. Sounders et al. (2009) differentiated between the two kinds of observation

in which participant observation is qualitative and stresses the discovery of meanings

that people attach to their actions. Structured observation however, according to

Sounders et al. (2009), is quantitative and is more concerned with the frequency of

those actions. Kothari (2004) explained that an observation is defined as information

collected by an investigator’s own observation without having interviewed the

respondents. Kothari (2004) also outlined the disadvantages of the observation method:

costs, the provision of limited information and its unsuitability for larger samples.

68

For the proposed research project, an observation method is used that has a direct

understanding of the process by which fundus images involving diabetes mellitus

patients are captured using a fundus camera. It also includes the manual grading process

of the captured fundus images by trained health care staff during the screenings, as

shown in Figure 3.3. The observation is important in order to understand the following:

how the fundus camera works, the process of fundus images capture and the manual

grading of fundus images, which is aimed to be converted into a fully automated

grading process.

Figure 3.3 Fundus images capturing and the grading process

Visual acuity test Fundus camera KOWA VX-10

Fundus camera capture Grading process

This item has been removed due to Data Protection. The unabridged version of the

thesis can be found in the Lancester Library, Coventry University.

This item has been removed due to Data Protection. The unabridged version of the

thesis can be found in the Lancester Library, Coventry University.

69

3.5.2 Collecting Primary Data through Interview

According to Kothari (2004), a personal interview session requires the investigator to

follow a firm procedure and seek answers to a set of predetermined questions. In

additions, Kothari (2004) claimed that an interview is usually performed in an organised

manner and the output produced will be determined by the interviewer’s capability.

Face to face interviews with an ophthalmologist and a medical assistant involved in the

diabetic retinopathy screening are performed for the purposes of this study. A personal

interview with the ophthalmologist is required to gather valuable information about the

ophthalmology field, specifically on diabetic retinopathy, as well as the features of

diabetic retinopathy signs, the diabetic retinopathy screening process and the respective

management of patients with diabetic retinopathy problem. In addition, discussion with

respect to the ground truth from the collected data is held with more experienced

ophthalmologists. The ophthalmologists will determine the types of retinopathy for each

of the data examples based on the retinopathy signs found on the fundus images. The

produced ground truth is important with respect to the system testing phase, where the

manual diabetic retinopathy analysis by several experienced ophthalmologists will be

compared to the results of the automated system developed in this research project.

3.5.3 Collecting Secondary Data through Existing Fundus Images

The most important data for this research study are fundus images which are taken from

the fundus camera located in the screening room. The fundus images from the diabetic

patients are captured by experienced medical staff for screening. For the purposes of

this research, numerous clinical fundus retinal images from the personal computer

attached to the fundus camera are extracted and analysed. The developed system

proposed by this study is tested on both normal and colour fundus images obtained from

patients with diabetes mellitus. Manual diabetic retinopathy diagnosis, carried by an

experienced paramedic is, then compared to the results of the automated system in order

to test and improve the sensitivity and specificity of the proposed grading methods.

70

3.6 Data Management and Analysis Procedure The task of data analysis will take place after the data collection phase. This requires

several correlated operations, for example the categories establishment and their raw

data application such as coding and formulation, followed by statistical inferences

(Kothari, 2004).

For this research project, manual analysis by experienced paramedics is compared to the

results of the automated system based on the provided ground truth. The sensitivity and

specificity are calculated to test the capability of the proposed system and its potential

as a quality assurance in retinal screening. Sensitivity is the percentage of abnormal

fundus images which have been classified as abnormal, while specificity is the

percentage of normal fundus images classified as normal by the screening. Accuracy

can also be calculated as a feature of the screening’s quality assurance. According to the

UK National Institute for Clinical Excellence, well-designed screening studies are

required to determine whether new tests of screening or early detection (such as digital

camera retinal photography) meet the standards of 80% sensitivity and 95% specificity

(National Institute for Clinical Excellence, 2002). Listed below are four possible

outcomes of a screening. Results are required in sensitivity, specificity and accuracy

calculations (Taylor and Batey, 2012):

i. True negative: the image is normal and was reported by the screener as

normal

ii. True positive: the image shows retinopathy and was correctly reported by the

screener as having retinopathy

iii. False negative: the image shows retinopathy but was reported by the

screener as normal

iv. False positive: the image is normal but was reported by the screener as

showing retinopathy

Table 3.1 shows the formula to calculate the sensitivity, specificity and accuracy

respectively, as explained above.

72

conduct the research and collect medical data from Malaysian health clinics and

hospitals. Patient confidentiality and medical data protection are the important ethical

issues that need to be considered in this case.

3.8 Benchmark Public Data Sets There are several public data sets available for the benchmarking of diabetic retinopathy

detection from digital images. In order to run and test the developed system, firstly,

fundus images from publicly available data sets will be used as a benchmark before

running the system with the fundus images collected from the Melaka Hospital. Some

of the most popular public databases which contain eye fundus images are the Standard

Diabetic Retinopathy Database Calibration Level 0 (DIARETDB0), the Standard

Diabetic Retinopathy Database Calibration Level 1 (DIARETDB1), Methods to

Evaluate Segmentation and Indexing techniques in the field of Retinal Ophthalmology

(MESSIDOR), Digital Retinal Images for Vessel Extraction (DRIVE), STructured

Analysis of the Retina (STARE), the Retinal Vessel Image set for Estimation of Widths

(REVIEW) and the Retinopathy Online Challenge (ROC) database.

One of the more popular data sets containing a combination of normal and diabetic

retinopathy fundus images is the Standard Diabetic Retinopathy Database Calibration

Level 0 (DIARETDB0). The data set consists of 130 colour fundus images of which 20

are normal and the remaining 110 contain signs of diabetic retinopathy, such as hard

exudates, soft exudates, microaneurysms, haemorrhages and neovascularisation. The

original images, sized 1500 x 1152 in PNG format, are captured with a 50 degree field-

of-view digital fundus camera with an unknown camera setting (Kauppi et al., 2006). In

addition to the DIARETDB0, there is another data set developed by the Machine Vision

and Pattern Recognition Research Group at Lappeenranta University of Technology,

Finland, which is the Standard Diabetic Retinopathy Database Calibration Level 1

(DIARETDB1), with 89 colour fundus images, including 84 images with at least mild

non-proliferative signs (microaneurysms) of diabetic retinopathy in addition to five

normal images (Kauppi et al., 2007).

73

The methods to evaluate the segmentation and indexing techniques in the field of retinal

ophthalmology or the MESSIDOR database is another data set, produced by research

funded by the French Ministry of Research and Defence to facilitate studies on diabetic

retinopathy diagnosis (Messidor, 2004). It consists of 1,200 colour fundus images

captured using a colour video 3CCD camera on a Topcon TRC NW6 non-mydriatic

retinograph with a 45 degree field of view. The images acquired by three

ophthalmologic departments have three different sizes: 1440 x 960, 2240 x 1488 and

2304 x 1536 pixels and 8 bits colour representation (Messidor, 2004).

Another retinal images database is from the Digital Retinal Images for Vessel

Extraction (DRIVE) project, which offers both retinal colour images and the results of

the automatic segmentation of blood vessels. The set of 40 images, where 33 do not

show any sign of diabetic retinopathy and seven show signs of mild early diabetic

retinopathy, are captured using a Canon CR5 non-mydratic 3CCD camera with a 45

degree field of view, 8 bits per colour plane and a size of 768 by 584 pixels (Staal et al.,

2004).

The STructured Analysis of the Retina (STARE) project by Dr. Michael Goldbaum at

the University of California, San Diego, funded by the U.S. National Institutes of

Health, is another database with retinal colour images (Hoover et al., 2003). The set of

400 raw images including a list of diagnosis codes and a diagnosis for each image can

be obtained from the STARE database. Blood vessel segmentation work involves 40 of

these images (Hoover et al., 2000), while 80 images are used for optic nerve detection

(Hoover et al., 2003).

The DRIVE and STARE data set are excellent databases of retinal vessel pixel

segmentations; however they do not include width measurements. Therefore, the

Retinal Vessel Image set for Estimation of Widths (REVIEW) data set is presented to

fill this gap. The data set includes 16 images with 193 vessel segments and a variety of

pathologies and vessel types. The database contains accurate width measurements and

four subsets of images, which are categorised in four classes: high resolution, vascular

disease, central light reflex and kick-points. Al-Diri and others (2008) have described

74

the REVIEW data set for retinal vessel and the algorithm used to process the

segmentation in order to produce vessel profiles.

The Retinopathy Online Challenge (ROC) presents an online competition for numerous

methods for the detection of microaneurysms which can be compared using the same

data set (Niemeijer et al., 2010). The images have three different sizes: 768 x 576, 1058

x 1061 and 1389 x 1383. The data set consists of 50 training images of colour fundus

photographs with available reference standards, and 50 test images where the reference

standard was withheld by the organisers. The overall results show that the detection of

microaneurysms has been a challenging task for both automatic methods and human

expertise.

3.9 Developed Data Set In addition to the public data sets presented above, a combination of normal and

Diabetic Retinopathy (DR) fundus images from a novel data set was developed as part

of this research.

The fundus images are collected from the Eye Clinic, in the Department of

Ophthalmology, at the Melaka Hospital, Malaysia. The novel data set consists of 600

colour fundus images collected from 300 patient’s folders. Each of the patient’s folders

has a minimum of two images, at least one for the right side and one for the left side,

where two different angles were captured; with the optic disc and the macula

respectively at the centre. The original images, which are sized 3872 x 2592 in JPEG

format, provide high quality and detail. These were captured with a KOWA VX-10

digital fundus camera. Figure 3.4 shows some examples of the images from the new

developed data set. Three experts from the Department of Ophthalmology at the Melaka

Hospital, Malaysia were involved in order to diagnose the fundus images into ten

retinopathy stages: No Diabetic Retinopathy, Mild DR without maculopathy, Mild DR

with maculopathy, Moderate DR without maculopathy, Moderate DR with

maculopathy, Severe DR without maculopathy, Severe DR with maculopathy,

Proliferative DR without maculopathy, Proliferative DR with maculopathy and

Advanced Diabetic Eye Disease (ADED). An Excel file containing the link to each eye

75

fundus image and the retinopathy stages drop down list, as presented in Figure 3.5 was

provided to each of the experts separately in order to avoid bias. The summary findings

of the three experts are presented in Figure 3.6. The average from the three experts is

used for the overall expert diagnosis, as shown in Figure 3.7. As a result of the analysis

of the experts’ diagnosis, performed using SPSS software, the total number of images in

each class is as follows: normal (no retinopathy) class with 276 images, while the

abnormal or diabetic retinopathy (DR) class can be divided into nine other categories

including mild DR without maculopathy (72), mild DR with maculopathy (27),

moderate DR without maculopathy (85), moderate DR with maculopathy (83), severe

DR without maculopathy (23), severe DR with maculopathy (11), proliferative DR

without maculopathy (6), proliferative DR with maculopathy (10) and, finally, advanced

diabetic eye disease, ADED (7). These are presented in Table 3.2. Meanwhile, the data

analysis procedures of the novel dataset are discussed in Chapter 5 (Overall Results

Analysis and Discussion), together with the system results analysis.

Macula centre

Optic disc centre

Figure 3.4 Examples of images in the dataset

76

Figure 3.5 Expert diagnosis file

Figure 3.6 Expert diagnosis summary

77

Figure 3.7 Expert diagnosis average

Table 3.2 Expert diagnosis summary

Retinopathy Stage No. of Images

No DR 276

Mild DR without maculopathy 72

Mild DR with maculopathy 27

Moderate DR without maculopathy 85

Moderate DR with maculopathy 83

Severe DR without maculopathy 23

Severe DR with maculopathy 11

PDR without maculopathy 6

PDR with maculopathy 10

ADED 7

Total 600

The new dataset is different when compared to the other datasets, presented earlier in

Section 3.8. The data set represents the South East Asian population, particularly

Malaysian, unlike the other datasets, which represent the Caucasian population. It

provides an almost balanced total number of No DR/Normal and DR/Abnormal images.

Typically, in the diagnosis of medical images, it is difficult to find a large number of

normal cases. The imbalanced number of available images poses some problems in the

classification phase; therefore, the balanced number in the new dataset will help

overcome this problem.

78

Moreover, the categorisation of the expert diagnosis followed the standard practice

based on the International Clinical Retinopathy and Diabetic Macula Oedema Disease

Severity Scale. The classification of the data involves maculopathy, which is the yellow

lesion near the macula. This is a very detailed categorisation compared to other data

sets. The detection of maculopathy is very important, as the macula is responsible for

central vision and it represents a sensitive part of the eye. It is vital therefore in

detecting the urgency of a referral. In addition, the novel dataset and the expert

diagnosis may be used separately for both diabetic retinopathy grading and diabetic

maculopathy grading.

In order to make the novel fundus images dataset widely accessible, it has been made

available as an online database. The novel dataset is accessible at

http://creative.coventry.ac.uk/fundus. The webpage of this research contains the novel

dataset with eye fundus images, including the expert diagnosis file and the published

papers related to this research project. Figure 3.8 shows the screenshot of the main page

of the dataset webpage. The aim of this online database is to highlight the research

project development and to promote research on retinal-imaging to enable comparative

studies and, most importantly, to share the eye fundus images with other researchers.

The dataset can be downloaded for research and educational purposes.

Figure 3.8 Main web page for the developed dataset

79

3.10 Summary The chapter discussed the main phases involved in the research development, beginning

with the research design outline which implemented in the study. The data collection

and analysis are important stages in the research design. In this respect, the data

collection instruments were presented focusing on the data collection process design

and management. Ethical issues should also be considered in research, when dealing

with people and organisations. The chapter also provided information on the ethical

considerations involved in this research. Significantly, a high quality of data, i.e.,

fundus images and adequate techniques, can help rise the quality and efficiency of the

developed system, therefore, eye fundus images new dataset which would be beneficial

to researchers and practitioners in the area of retinal imaging was introduced, in addition

to the other public datasets used for the purposes of evaluation.

Table 3.3 shows a mapping table that lists the datasets used for each system

development in this research work. The Standard Diabetic Retinopathy Database

Calibration Level 0 (DIARETDB0) is used for the evaluation of System I, while the

Standard Diabetic Retinopathy Database Calibration Level 1 (DIARETDB1) is used for

System II evaluation. The Retinopathy Online Challenge (ROC) dataset is used for the

evaluation of microaneurysms detection in System III and System IV. The new

developed dataset collected from Melaka Hospital is used for the evaluation of System

VI and System VII, for the maculopathy and diabetic retinopathy detection. In addition

to the justification of the datasets used, Table 3.3 provides the type of the system

detection.

80

Table 3.3 Summary of datasets used for systems development

System I System II System III System IV System V System VI System VII

Dataset Used

Benchmark Public Dataset:

- DIARETDB0 √

- DIARETDB1 √

- MESSIDOR

- DRIVE

- STARE

- REVIEW

- ROC √ √

Developed Dataset:

- Melaka Hospital √ √ √

Type of System

Development

General detection General detection Microaneurysms

detection

Microaneurysms

detection

Maculopathy

detection

Maculopathy

detection

Maculopathy

detection

Justification of Selected

Dataset

- A combination

of normal and

diabetic

retinopathy

fundus images

- Suitable for

general

classification

- Contains

normal fundus

images and

images with

mild NPDR

- Suitable for

general

classification

- ROC presents an online competition

for numerous methods in

microaneurysms detection to

compare with each other on the same

data

- Suitable for the microaneurysms

detection and classification

- New dataset consists of 600 fundus images and experts

diagnosis which diagnose the fundus images into ten

retinopathy stages, following the ophthalmologists

practice

- Suitable for the maculopathy and diabetic retinopathy

detection and classification

81

4 CONTRIBUTIONS TO DEVELOPING SYSTEMS FOR

DIABETIC RETINOPATHY AND MACULOPATHY

DETECTION

Several systems, implementing different methods, have been proposed for diabetic

retinopathy screening in the course of this research. The development of the proposed

systems can be categorised into three types of approach. The first system development

is based on a general detection of diabetic retinopathy (i.e., either normal or with

retinopathy present). Meanwhile, the second type presents several system developments

for the detection of microaneurysms, an important sign of diabetic retinopathy. Finally,

the third type explains the development of two systems in order to detect diabetic

retinopathy alongside maculopathy.

This chapter presents the development and evaluation of each of the proposed types of

system development. They present a combination of techniques, such as different

preprocessing techniques, different feature parameters and different classifiers in a

diabetic retinopathy screening system, which differs from systems proposed by other

researchers. The preliminary screening system (System I) for the general detection of

diabetic retinopathy is presented in Section 4.1. The detection of microaneurysms in

colour fundus images is presented in Section 4.2. The first two systems (System II and

System III) for the detection of microaneurysms, which highlight the use of feature

extraction and classification methods, are explained in Section 4.2.1 and Section 4.2.2,

respectively. The other two proposed systems (System IV and System V) for the

microaneurysms detection, which implement the fuzzy image processing techniques, are

reported in Section 4.2.3 and Section 4.2.4. Section 4.3 presents the two systems

(System VI and System VII) for the detection of diabetic retinopathy alongside the

detection of maculopathy in the eye’s fundus images. Finally, a summary of the chapter

is presented in Section 4.4.

82

4.1 Basic System for General Diabetic Retinopathy Detection

(System I) This section presents a preliminary system for the classification and screening of

diabetic retinopathy using eye fundus images. This is a general detection system of

diabetic retinopathy that classifies images into two main classes: normal and with

diabetic retinopathy. The system is evaluated by using the Standard Diabetic

Retinopathy Database Calibration Level 0 (DIARETDB0), described in Section 3.8,

which consists of 130 colour fundus images (Kauppi et al., 2006).

The system explored and implemented some basic image processing techniques, which

will be used for further developments of the diabetic retinopathy screening system.

Rahim and others (2014) reported the development of such an automatic screening and

classification of diabetic retinopathy fundus images in detail, while exploring the

existing systems and applications related to diabetic retinopathy screening and detection

methods. This system is an automatic system for detecting diabetic retinopathy by

classifying the images into general detection categories, which are normal (no apparent

retinopathy) or abnormal (retinopathy present).

The proposed ophthalmic decision support system consists of an automatic acquisition,

screening and classification of diabetic retinopathy fundus images, which will assist in

the detection and management of diabetic retinopathy. The developed system contains

four main parts, namely the image acquisition, the image preprocessing, the feature

extraction, and the classification using several machine learning techniques.

The proposed screening system has been developed using open source software,

OpenCV (Open Source Computer Vision) and Microsoft Visual C++ 2010. The

OpenCV environment, developed by Willow Garage, is a programming library offered

for real time computer vision (Itseez, 2014). OpenCV includes a collection of

standardised image analysis and machine vision algorithms to be used by developers.

Most work in this area uses tools such as Matlab and SPSS for feature extraction and

analysis, but by using OpenCV it is possible to build more effective systems, with

processing times that are suitable for use in real situations. Using OpenCV also

simplifies the distribution of software due to permissive licensing, and it lowers the cost

83

of development, use and maintenance because there are no purchases or licensing fees.

Finally, OpenCV is portable, meaning that any machine that is able to run C is also

likely to be able to run OpenCV. Furthermore, OpenCV has been used on Windows,

Linux, MacOS and Android systems.

The proposed system starts with the image acquisition process, where images are

selected for further processing. These will undergo preprocessing in order to improve

the image contrast in addition to other enhancements. The preprocessed images will

then be used to extract a number of features, such as the area, the mean and the standard

deviation of on pixels. Four nonlinear classifiers, namely a binary decision tree, a k-

nearest neighbour classifier, and two support vector machines, using radial basis

function and polynomial function kernels respectively, will then be trained on the

training set to find an optimal way to group images into their respective classes. Finally,

in the prediction phase, where the system might ultimately be used to assist the

clinician, the images are classified into two main cases: normal or diabetic retinopathy.

Figure 4.1 presents the block diagram of the proposed system for automating the

screening and classification of diabetic retinopathy.

Figure 4.1 Block diagram of the proposed general automatic screening and

classification of diabetic retinopathy

84

4.1.1 Image Preprocessing

Preprocessing is the process of image data improvement, involving the enhancement of

some image characteristics or features for the next stage of processing. The image

preprocessing techniques involved in the present work include Greyscale Conversion,

Contrast Limited Adaptive Histogram Equalisation, Discrete Wavelet Transform,

Filtering and Morphological Operations.

The first preprocessing technique used is the conversion of the colour fundus image into

a greyscale image, as greyscale is usually a better format for image processing. A

greyscale image has pixels of a single value, namely its intensity information. It is also

known as a “black and white” image. The intensity is calculated by using a common

formula combination of 30% of red, 59% of green and 11% of blue.

The Adaptive Histogram Equalisation (AHE) is a computer image processing technique

for improving the image’s contrast. The difference between the adaptive histogram

equalisation and the ordinary histogram equalisation is that the adaptive histogram

equalisation computes several histograms for different sections of the image, and

subsequently distributes the lightness values. This technique is used to improve local

contrast and enhance more details of the image. However, the adaptive histogram

equalisation has its limitations, as it produces over-amplification of noise in the

homogeneous regions of an image. Therefore, the Contrast Limited Adaptive Histogram

Equalisation (CLAHE) is used in the proposed system in order to prevent the over

amplification of noise. CLAHE functions by clipping the histogram at the predefined

value before computing the cumulative distribution function. Histogram equalisation

processed image is obtained by mapping each pixel with level rk in the input image into

a corresponding pixel with level sk in the output image, as given below (Gonzalez and

Woods, 2002):

85

𝑠𝑘 = 𝑇(𝑟𝑘) = ∑ 𝑝𝑟

𝑘

𝑗=0

(𝑟𝑗)

= ∑𝑛𝑗

𝑛

𝑘

𝑗=0

𝑘 = 0, 1, 2, … , 𝐿 − 1

(4.1)

Discrete wavelet transform is a variant of wavelet transform for which the wavelets are

discretely sampled. The discrete wavelet transform is an O(N) algorithm and it is also

often referred to as the fast wavelet transform. The Haar wavelet is implemented in the

proposed system development as it is a simple wavelet transform and it is currently used

in many methods of discrete image transforms and processing. Discrete wavelet

transforms can be used to reduce the image size without losing much of the resolution.

The implementation of discrete wavelet transform is necessary to overcome the

limitation of the software used, where a warning is generated when the image is too big

to fit on the screen. Therefore resizing of the image is required. Since the fundus images

are of a high resolution and of quite a large size, the use of the Haar wavelet is

recommended. As a result of the implementation of the Haar Wavelet on the proposed

system, the size of the fundus images are reduced by half, from 1500x1152 to 750x576.

Image filtering is used to improve the image quality or to restore a digital image which

has been corrupted by some noise. A comparison of the performance between three

different edge operators, i.e., Sobel, Prewitt and Kirsch has been proposed for the

detection and segmentation of blood vessels in the colour retinal images (Karasulu,

2012). The experimental results show that the edge-based segmentation using the

Kirsch compass templates is far superior to other methods. Moreover, the Kirsch

operator can adjust the related threshold value automatically due to the image

characteristics. Based on these reasons, the Kirsch operator has been chosen as an edge

detection filter technique in the proposed system development. The Kirsch edge

detection uses eight filters (i.e., eight masks for the related eight main directions) that

are applied to a given image in order to detect edges. These eight filters are a rotation of

a basic 3x3 compass convolution filter (i.e., single mask). The Kirsch filter is applied on

86

the wavelet transform image to create the eight filtered output images. The masks are

distinct as given below (Karasulu, 2012):

𝑀0 = [5 5 5

−3 0 −3−3 −3 −3

] , 𝑀1 = [5 5 −35 0 −3

−3 −3 −3] , 𝑀2 = [

5 −3 −35 0 −35 −3 −3

],

𝑀3 = [−3 −3 −35 0 −35 5 −3

] , 𝑀4 = [−3 −3 −3−3 0 −35 5 5

] , 𝑀5 = [−3 −3 −3−3 0 5−3 5 5

],

𝑀6 = [−3 −3 5−3 0 5−3 −3 5

] , 𝑀7 = [−3 5 5−3 0 5−3 −3 −3

]

(4.2)

Morphological operations are used for certain purposes including image preprocessing,

enhancing object structure, segmenting objects from the background and also for the

quantitative description of objects (Sonka et al., 2008). In the proposed system

development, morphology operators involving dilation and erosion are implemented to

extract the blood vessels. A closing operation is defined as dilation followed by the

erosion operator. Joshi and Karule (2012) implemented the closing operation for retinal

blood vessel segmentation, where the disk shaped structuring element for the

morphological operation is used. The dilation operates in greyscale images to enlarge

brighter regions and it closes the small dark regions, while the erosion operator shrinks

the dilated objects back to their original size and shape. As a result, the vessels being

thin dark segments laid out on a brighter background, are closed by the closing

operation. Figure 4.2 (a)-(f) shows the output after each of the preprocessing operations

on a selected image.

87

(a) Original image (b) Greyscale

Conversion

(c) Adaptive

Histogram

Equalisation

(d) Discrete Wavelet

Transform (e) Kirsch Filtering

(f) Morphological

Operators

Figure 4.2 Image preprocessing

4.1.2 Feature Extraction

After performing the preprocessing techniques, feature extraction takes place in order to

obtain the features from the given images. Features can be grouped into shape-based

features, pixel-intensity-based features, Fourier-descriptor-based features and colour-

based features. Within the preliminary system for the classification and screening of

diabetic retinopathy, three basic shape features, including the area of on pixels and the

mean and the standard deviation, are extracted for the purpose of diabetic retinopathy

detection. The three values have been chosen as they are basic features and suitable for

the pre-processed candidate image. These values for both normal and diabetic

retinopathy images are used in order to create a model for training. Table 4.1 presents

the details of the feature extracted including the generated code.

88

Table 4.1 Feature extraction in the proposed system

4.1.3 Classification

The extracted feature values from the step described above have been passed to the

classification stage. The PRTools, a Matlab toolbox for pattern recognition has been

used to implement the classifiers (Duin et al., 2007). Nonlinear classifiers can provide

better classification results compared to linear classifiers. Therefore, four nonlinear

classifiers, namely the binary decision tree classifier, the k-nearest neighbour classifier,

the radial basis function (RBF) kernel based support vector classifier and the

polynomial kernel based support vector classifier have been selected to be trained on

images and to group them into two classes, i.e., normal and diabetic retinopathy

respectively, based on the three extracted features. The decision tree is a classifier in the

form of a tree structure. It classifies instances by starting at the root of the tree and

moving through it until a leaf node is reached. In the k-nearest neighbour classifier, the

object is classified by a ‘majority vote’ of its neighbours, with the object being assigned

to the most common class among its k nearest neighbours. The 1-nearest neighbour rule

(1-NN) is used in the particular implementation of the system. A support vector

machine (SVM) performs the classification by constructing an N-dimensional

hyperplane that optimally separates the data into two categories. The support vector

machine classifier can use various kernel functions, such as linear, polynomial or a

radial basis function. The kernel function transforms the data into a higher dimensional

space in order to be able to perform the separation in the nonlinear region. Two different

Feature Description Snippet Code

Area of

on pixels

Number of grey level pixels on

the black and white image, where

white pixels on are all pixels

above a threshold of 100 pixels

maxS = cvGet2D(gray, y, x);

val = maxS.val[0];

if(val > 100) {

count++;

sum += val; }

Mean Mean value of on pixels mean = sum / count;

Standard

deviation

Standard deviation of on pixels

maxS = cvGet2D(gray, y, x);

val = maxS.val[0];

if(val > 100) {

count++;

val = (mean-val);

sum += val*val; }

sdv = sqrt(sum) / count;

89

types of kernel functions provided in Matlab for SVM classification were used, i.e., the

second order polynomial kernel SVM, svc (ATrain, ‘p’, 2), and the radial basis function

kernel SVM, rbsvc (ATrain). The results show that the RBF kernel outperformed the

results obtained by the second order polynomial kernel.

4.1.4 System Results and Evaluation

Figure 4.3 shows the user interface snapshot of the proposed developed system. The

performance (misclassification error) of the four classifiers is presented in Table 4.2.

Since the dataset is hugely unbalanced, the minority class was oversampled by

duplication in order to balance it. The DIARETDB0 data is split randomly into 90% for

training and the remaining 10% for testing. The process is repeated ten times in a cross-

validation procedure in order to generate unbiased results. The average results on the ten

runs for each of the four classifiers are then reported. The four machine learning

classifiers were chosen to show the variance performance of different categories of

algorithms in machine learning, i.e., clustering, classification and regression. For

example, for the classification category, Support Vector Machines and k-nearest

neighbour were used, and for the regression category, the binary decision tree was used.

For more clarity, in Table 4.2, the confusion matrix for the first out of the ten

experiments is presented, in order to show the relative performance of the four

classifiers. A total of 130 colour fundus images, which consists of 110 normal images

and 20 diabetic retinopathy images, were used initially. Later, the diabetic retinopathy

images were duplicated to contribute the same number as the normal images. As a

result, both normal and diabetic retinopathy classes constitute a total of 220 images,

where each class has 110 images, respectively. A 10% from the 220 images, which total

to 22 images, were used for the testing part, dividing 11 images for the normal class and

another 11 images for the diabetic retinopathy class. The values of true positives, false

negatives, false positives and true negatives can be extracted from the generated

confusion matrix for the calculation of specificity, sensitivity and accuracy of the

system. For example, for the binary decision tree, the value of true positives is 6, false

negatives is 5, false positives is 0 and the true negatives is 11. The classification

performance of the diagnosis system is assessed using the accuracy of the individual

classifiers and also their specificity and sensitivity. The experimental results show that

90

the four classifiers, and especially the k-nearest neighbour, are well able to identify both

classes, i.e., the normal and the diabetic retinopathy (DR) cases. All four classifiers had

more success in identifying the diabetic retinopathy cases, as there were more examples

of such images in the database. As a comparison, the system results proposed by Priya

and Aruna (2011) showed a sensitivity of 99.45%, specificity of 100% and accuracy of

98.92% for the Support Vector Machine classifier. In addition, the experimental results

in (Priya and Aruna, 2012) showed that the results were 98%, 96% and 97.6% for

sensitivity, specificity and accuracy, respectively, for Support Vector Machine, while

for Probabilistic Neural Network the results were 90% for sensitivity, 88% for

specificity and 89.6% for accuracy.

Figure 4.3 Snapshot of the proposed system user interface

Table 4.2 Average results when using the four classifiers

Binary decision

tree

k-nearest

neighbour

RBF kernel

SVM

Polynomial kernel

SVM

Misclassification

error

0.2091 0.01364 0.0909 0.3182

Accuracy 0.7909 0.9864 0.9091 0.6818

Specificity 1 1 1 0.5545

Sensitivity 0.5818 0.9727 0.8182 0.8091

Confusion matrix for

the first experiment

Labels

(1 : Normal, 2: DR)

True | Estimated Labels

Labels | 1 2 | Totals

-----|----------|--------- 1 | 11 0 | 11

2 | 5 6 | 11

----|----------|--------- Totals | 16 6 | 22



----|----------|--------- 1 | 11 0 | 11

2 | 0 11 | 11

------|-----------|--------- Totals | 11 11 | 22



-----|-----------|--------- 1 | 11 0 | 11

2 | 2 9 | 11

------|----------|--------- Totals | 13 9 | 22



------|---------|--------- 1 | 6 5 | 11

2 | 2 9 | 11

-------|---------|--------- Totals | 8 14 | 22

91

A general automatic system for screening and classification of diabetic retinopathy

using fundus images has been developed. The classification area of the system can be

enhanced by building an ensemble of classifiers. Unbalanced learning techniques can

also be considered in order to train the individual classifiers in the ensemble. In

addition, more sophisticated features will be used in the future work to properly

discriminate between the various diabetic retinopathy signs (i.e., different features

extraction for microaneurysms, haemorrhages, exudates, etc.). The system will also be

extended to obtain more details of the diabetic retinopathy classification, namely to

classify them into the following cases: no apparent retinopathy, mild non-proliferative,

moderate non-proliferative, severe non-proliferative and proliferative DR. In addition to

the classification diagnosis, the system can be improved to provide the recommended

follow-up schedule for each stage, as underlined by the American Academy of

Ophthalmology, in order to become a complete system to be used in a diabetic

retinopathy screening practice.

4.2 Microaneurysms Detection Microaneurysms, which are the earliest visible sign of diabetic retinopathy, appear as

small dots on the retina. The early detection of microaneurysms is the first step in

preventing diabetic retinopathy. Since the detection of microaneurysms is important and

challenging, four variants of system development (System II – System V) which

introduce different configurations and different techniques, are presented in this chapter

(see also Rahim et al., 2015a, 2015b).

4.2.1 Automatic Detection of Microaneurysms in Colour Fundus

Images Using Vessel Segmentation and Features Extraction

(System II)

The system for the automatic detection of microaneurysms presented in this section

consists of four main parts, namely, the image acquisition, the image preprocessing, the

feature extraction, and the classification, using several machine learning techniques.

System II is evaluated using the Standard Diabetic Retinopathy Database Calibration

Level 1 (DIARETDB1), containing 89 colour fundus images (Kauppi et al., 2007).

92

There are 75 images with microaneurysms and the remaining 14 images were identified

as showing no signs of microaneurysms.

The initial stage of the proposed system is the image acquisition process, followed by

the preprocessing process. The preprocessed images are then used to extract a number

of features. Four nonlinear classifiers, namely a binary decision tree, a k-nearest

neighbour classifier, and two support vector machines, using radial basis function and

polynomial function kernels respectively, are then trained on the training data to find an

optimal way to group images into their respective classes. Finally, in the classification

phase, the images are classified as to whether microaneurysms are present or not. The

overall process of microaneurysm detection is shown in Figure 4.4.

Figure 4.4 Block diagram of the proposed automatic detection of

microaneurysms

93

4.2.1.1 Image Preprocessing

Preprocessing is used for image improvement. Greyscale Conversion and Shade

Correction are the preprocessing techniques used in the proposed system.

Firstly, the colour fundus image is converted into the greyscale format for improved

contrast. The second technique is shade correction, where the background image is

estimated and later subtracted from the original image. The non-uniform illumination in

the image needs to be corrected if the presence of microaneurysms in this area is to be

correctly detected. The first step is to estimate the background using the morphological

opening technique. The function imopen in Matlab (MathWorks, 2014) is used to

perform a morphological opening of the greyscale image (with the structuring element

of 12 pixels and disk shape). The second step is to subtract the background image from

the original image (using the imsubtract function). The shade correction process is then

continued, by increasing the image contrast (using the function imadjust), followed by

thresholding of the image (by using im2bw function), and finally by the removal of the

background noise (using bwareaopen). As a result, the thresholded image is inverted

and the final output shows only the fundus image area.

After the greyscale conversion and shade correction, vessel segmentation is then

performed. The vessels are extracted from the shade corrected image using a

morphological operation. The image is closed using a disc shaped structuring element of

5 pixels. The shade corrected image is filled to eliminate holes in the vessels. Later, the

filled image is subtracted from the closed image to give a vessel difference image. The

image is thresholded to obtain binary images containing the vessels. The binary image

is subtracted from the Gaussian filtered image, so that the final image has vessel free

candidates.

In addition to the two main preprocessing methods of greyscale conversion and shade

correction, other image preprocessing techniques can also be implemented, such as the

green channel conversion, the median filter, the Gaussian filter and the Contrast Limited

Adaptive Histogram Equalisation. In addition, four of the blood vessel extraction

techniques can be implemented, including the Kirsch Template, the Frangi Filter, Local

94

Entropy and Entropic Thresholding. Figure 4.5 shows the output after the preprocessing

operations are performed on a selected image.

Figure 4.5 Preprocessing the output image

4.2.1.2 Feature Extraction

After performing the preprocessing tasks, feature extraction takes place in order to

obtain relevant features from the given images. Since a microaneurysm has specific

features, such as a circular shape and a red colour, appropriate features should be

extracted to ensure reliable feature extraction and classification. Sopharak and

colleagues (2013) have listed some useful features for microaneurysm detection based

on shape, pixel intensity, Fourier descriptor and colour. Features for the purposes of

microaneurysm detection, such as the area of the pixels, the perimeter of the object, the

major and minor axis length, its aspect ratio and its circularity have been chosen and

extracted in the second system. These values are used to extract microaneurysms as they

are appropriate for the characteristics of the microaneurysms. Table 4.3 presents the

details of the features extracted in our system.

(a) Greyscale Conversion (b) Shade Correction (c) Vessel Segmentation

95

Table 4.3 Feature extraction in the proposed microaneurysms detection system

Feature Description

Area of pixels in the

candidate object

The actual number of pixels in the region

Perimeter of the object The distance around the boundary of the region

Major Axis Length of the

candidate object

The length (in pixels) of the major axis of the ellipse that

has the same normalised second central moments as the

region

Minor Axis Length of the

candidate object

The length (in pixels) of the minor axis of the ellipse that

has the same normalised second central moments as the

region

Aspect ratio Major Axis Length divided by the Minor Axis Length

Circularity Roundness of the candidate, ((4*pi*Area) / (Perimeter. ^2))

4.2.1.3 Classification

Classification was carried out using the PRTools package (Duin et al., 2007) in Matlab.

In the second system, presented in this section, the classifiers selected and implemented

for image classification purposes were: the binary decision tree classifier, the 1-nearest

neighbour rule (1-NN) classifier, the radial basis function kernel based support vector

classifier and the second order polynomial kernel based support vector classifier.

4.2.1.4 System Results and Evaluation


performance (misclassification error) of the four classifiers is presented in Table 4.4.

Since the dataset is imbalanced, containing 75 images with signs of microaneurysms

and only 14 normal images, the minority class was therefore oversampled by one-time

duplication in order to avoid having an overly imbalanced dataset. The DIARETDB1

data is split randomly into 90% for training and the remaining 10% for testing. The

process is repeated ten times in a cross-validation procedure in order to generate

unbiased results. The average results on the ten runs for each of the four classifiers are

then reported.

The accuracy, sensitivity and also the specificity of the individual classifiers are

presented in Table 4.4 to measure classification performance. The accuracy of the four

classifiers i.e., the binary decision tree and the 1-nearest neighbour is 0.9091, while the

96

radial basis function kernel and the second order polynomial kernel based support

vector classifier is 0.7273. The experimental results show that the four classifiers are

able to identify both classes, i.e., the “microaneurysms detected” and the “no

microaneurysms” classes. The binary decision tree and the k-nearest neighbour

classifiers yielded very good results.



Binary

decision tree

k-nearest

neighbour

RBF kernel

SVM

Polynomial kernel

SVM

Misclassification error 0.0909 0.0909 0.2727 0.2727

Accuracy 0.9091 0.9091 0.7273 0.7273

Specificity 1 1 0.3333 0

Sensitivity 0.875 0.875 0.875 1

97


Images Using a Combination of Image Preprocessing Techniques

and Circular Hough Transform (System III)

The third system proposed here is an automatic detection of microaneurysms, which

introduces the combination of the preprocessing techniques and the Circular Hough

Transform for the localisation and detection of microaneurysms in colour fundus

images. For the third proposed systems, 40 fundus images of three different sizes (768 x

576, 1058 x 1061 and 1389 x 1383) have been used for evaluation from the Retinopathy

Online Challenge (ROC) public database (Niemeijer et al., 2010).

The Circular Hough Transform technique has been previously proposed for the

detection of microaneurysms in two types of photography modes, i.e., in retinal

fluorescein angiographic images (Abdelazeem, 2002) and also in digital red-free

photographs (Hipwell et al., 2000). The application of the Circular Hough Transform

for the detection of microaneurysms in the colour mode (another type of fundus

photography mode) was therefore explored. In the colour mode, the retina is illuminated

by white light and examined in full colour. The red-free or monochromatic mode is

where the imaging light is filtered to remove red colours, which helps to improve the

contrast of vessels and other structures. In the angiography mode, vessels are brought

into high contrast by the intravenous injection of a fluorescent dye, which produces a

very high-contrast image of the vessels. The detection of microaneurysms in colour

fundus images is more challenging compared to angiography and red-free types.

The initial stage of the proposed system is the image acquisition process, followed by

the preprocessing process. Following this, the detection of the microaneurysm is

performed using the Circular Hough Transform method. Finally, in order to test the

accuracy of the microaneurysms detection system compared to the ROC annotation,

statistical tests are performed for results analysis. The overall process of microaneurysm

detection is presented in Figure 4.7.

98

Figure 4.7 Block diagram of the proposed automatic detection of

microaneurysms using Circular Hough Transform


A combination of preprocessing techniques, i.e., a Greyscale Conversion and Contrast

Limited Adaptive Histogram Equalisation are used in the system development. The

proposed preprocessing techniques are used to help increase the contrast of the colour

fundus images. The uses of colour fundus images are more challenging compared to

other modes of fundus photography examination: angiography and red-free.

Appropriate techniques must therefore be implemented in order to improve the contrast

of the fundus images for improved visualisation and detection.

4.2.2.2 Circular Hough Transform

The Circular Hough Transform (CHT) is implemented in the proposed system to locate

microaneurysms, due to their circular shape. In addition, CHT is also useful in detecting

the optic disc in a diabetic retinopathy screening system. The Hough Transform has

been used for optic disc localisation by Noronha and other colleagues (2006).

99

Meanwhile, the same method was used to detect the early signs of diabetic retinopathy,

represented by microaneurysms (Abdelazeem, 2002).

The Hough Transform can be used to detect lines, circles or other parametric curves. It

can be used to determine a circle when a number of known points fall on the perimeter.

The objective is to find the (a, b) coordinates of the centres at (x, y) on a circle of radius

R. The Hough Transform offers some advantages, such as simplicity, ease of use, the

effective handling of missing and occluded data and the fact that it can be adapted to

many types of forms, other than lines. However, the limitations of the Hough Transform

include the complex computation involved for objects with many parameters, the

detection of one single type of object, difficulty in determining the length and position

of a line segment and finally, that it cannot separate collinear line segments (Solberg,

2009).

Firstly, the radius of the microaneurysm in the fundus images is calculated (using the

function imdistline). Based on the radius range specified, the system will find the circles

in the fundus images using CHT (through the imfindcircles function). Finally, after

finding the circles in the image based on the radius range, a circle is then created on the

current axes (using the function viscircles). The output of the third proposed system in

the localisation and detection of microaneurysms is shown in Figure 4.8.

Figure 4.8 Microaneurysms detection by using the Circular Hough Transform

100



performance analysis of the system and the expert annotation are presented in Table 4.5

and Table 4.6, respectively. The number of microaneurysms detected by the expert and

the developed system are calculated and compared. The presence of any number of

microaneurysms is represented as diabetic retinopathy (microaneurysms detected) and if

there is no microaneurysm detected, it is considered as normal (microaneurysm not

detected). Two types of statistical tests were performed to test the system performance

compared to the expert diagnosis. The T-test is used to test the mean differences

between the annotated images and the system output, and the results are presented in

Table 4.5. The fundus images (forty in total) diagnosed by the expert and by the system

were then compared with the T-test analysis. Firstly, a descriptive statistical analysis

was conducted and the results, as shown in Table 4.5, indicate that there is only a small

difference in the mean for the annotated images and that of the system (1.78 and 1.88,

respectively). The inferential statistical analysis (i.e., T-test) result indicates a p-value of

0.253. This shows that the means of the annotated images and the system output are not

significantly different.

Figure 4.9 Snapshot of the user interface of System III

101

Table 4.5 T-test analysis

*n=40

Mean

Standard

Deviation

Standard

Error

Mean p-value

Pair Tested 0.253

Expert 1.78 0.423 0.067

System 1.88 0.335 0.053

*n=Number of fundus images diagnosed by the expert and the system

The Chi-square test was also used to compare the two groups (i.e., the expert diagnosis

and the system diagnosis). In addition, the Chi-square test was used to explore the

relationship between these two categorical variables, as the results will help determine

whether the system findings are more likely to be different from the expert findings, or

in this case whether or not the expert is more likely to successfully identify diabetic

retinopathy than is the system. The results generated by the Chi-square test are shown in

Table 4.6, where the Chi-square test indicates a p-value of 0.239. Thus, it can be

concluded that the system findings are more likely to be no different to the expert

findings. In addition, Table 4.6 also shows the results of a cross tabulation generated to

descriptively compare the methods of assessment between two groups: expert and

system diagnosis for both normal and diabetic retinopathy categories. The results show

the expert diagnosed 9 images as normal and 31 images as having diabetic retinopathy,

while the developed system diagnosed 5 images as normal and the remaining 35 images

as belonging to the diabetic retinopathy category, providing a further indication of the

similarities between the two sets of diagnosis.

Table 4.6 Chi-square analysis

*n=80

Classification

Normal

n (%)

DR

n (%)

Total

n

Chi-Square Test Statistic

Method of

Assessment

χ2 (1, n=80) = 1.385,

p=0.239

Expert 9 (22.5) 31 (77.5) 40

System 5 (12.5) 35 (87.5) 40

*n=Number of fundus images diagnosed by the expert and the system

102


Images Using Fuzzy Image Processing (System IV)

The fourth system applies a fuzzy technique for image preprocessing in the detection of

microaneurysms. As explained in Section 4.2.2.1 above for the third system, the colour

fundus images are more challenging compared to the other modes of fundus

photography examination. Therefore, in order to obtain better visualisation and accurate

detection, contrast enhancement should be implemented.

The fourth system proposes the implementation of a fuzzy preprocessing technique, the

Fuzzy Histogram Equalisation. The fuzzy preprocessing techniques were used for

contrast enhancement in the medical digital images, such as pathology images (Garud et

al., 2011) and also other non-medical images (Sheet et al., 2010). The performance of

the fuzzy preprocessing techniques reported in previous work is promising. Therefore,

the suitability of the fuzzy preprocessing technique for the screening of diabetic

retinopathy using colour fundus images will be investigated and proposed.

Contrast enhancement produces a better image than the original by changing the pixel

intensities (Sheet et al., 2010). There are several contrast enhancement techniques

available: Histogram Equalisation (HE), Contrast Limited Adaptive Histogram

Equalisation (CLAHE), Histogram Stretching and brightness preserving histogram

modification approaches. Histogram equalisation is a technique for adjusting image

intensities in order to enhance contrast, in which the grey-level values are uniformly

distributed. Adaptive histogram equalisation is a more advanced version of histogram

equalisation which divides the image into smaller tiles, applies the histogram

equalisation to each tile, and then interpolates the results. This adaptive histogram

equalisation includes limits on how far the contrast should be changed, namely the

Contrast-Limited Adaptive Histogram Equalisation, CLAHE.

Sheet et al. (2010) proposed a modified technique of the brightness preserving

equalisation called the Brightness Preserving Dynamic Fuzzy Histogram Equalisation

(BPDFHE). This technique is explained at length in Section 2.7. The representation and

processing of images in the fuzzy domain allows the technique to handle the inexactness

of grey level values in a better way in order to improve the overall performance. The

103

technique proposed by Sheet et al. (2010) is used for contrast enhancements in digital

pathology images (Garud et al., 2011). The performance of this technique has been

compared with HE and CLAHE and, as a result, the BPDFHE preserved the image

brightness better than the other two techniques. Good performance of the BPDFHE

technique especially in medical images such as pathology images is reported by Garud

et al. (2011). This technique has been explored as a preprocessing technique for the

proposed detection of microaneurysms in diabetic retinopathy screening.

The first stage of the proposed system is the image acquisition process, the second is

preprocessing process, the third is the detection of microaneurysms using the Circular

Hough Transform method and, finally, in order to test the accuracy of the

microaneurysms detection system with the ROC annotation, the statistical tests are used

for analysis. System IV(a) is similar to the third system presented in Section 4.2.2,

which implemented the Greyscale Conversion, Contrast Limited Adaptive Histogram

Equalisation and Circular Hough Transform. The difference between these systems is

that the third system bases its categorisation of Normal (no microaneurysms detected)

or Diabetic Retinopathy (microaneurysms detected) on the presence of microaneurysms.

System IV on the other hand, categories results into microaneurysms detected or

undetected based on the number of microaneurysms detected between the expert and the

developed system. The overall process of microaneurysms detection is presented in

Figure 4.10.


A combination of preprocessing techniques including a Greyscale Conversion and a

Contrast Limited Adaptive Histogram Equalisation (CLAHE) are implemented for the

first system, System IV(a), while for the second system, System IV(b), a combination of

both a Greyscale Conversion and a Fuzzy Histogram Equalisation is proposed. Figure

4.11 shows both the output and the histogram for the intensity of the fundus image after

the preprocessing techniques have been applied, i.e., Greyscale Conversion, CLAHE

and Fuzzy Histogram Equalisation.

104

Figure 4.10 Block diagram of System IV for the automatic detection of

microaneurysms


(a) Greyscale

Conversion

(b) Contrast Limited

Adaptive Histogram

Equalisation

(c) Fuzzy Histogram

Equalisation

105

4.2.3.2 Microaneurysms Detection

The Circular Hough Transform technique is implemented in the fourth system to detect

the microaneurysms, as this technique has a good ability to detect circular shapes. The

background on the Circular Hough Transform technique is presented in Section 4.2.2.2.

4.2.3.3 System Results

Figure 4.12 shows the user interface snapshot of the proposed developed system.

Generally, the interfaces for both systems are similar, but the implementation of the

preprocessing techniques is different, as mentioned in Section 4.2.3.1 above.

The statistical result comparisons for both systems are presented in Table 4.7. The T-

test and the ANOVA test were performed to test the mean differences between the

annotated images and the system results based on the number of microaneurysms

detected by the expert and by the system. The first system, using a combination of

Greyscale Conversion and Contrast Limited Adaptive Histogram Equalisation, shows

that the means of the annotated images (7.48) and the system output (18.98) are

significantly different. On the other hand, the second system, which implemented the

Greyscale Conversion and the Fuzzy Histogram Equalisation, shows that the means of

the annotated images (7.48) and the system (9.33) are not significantly different. The

inferential statistical analysis (i.e., T-test) result indicates a p-value of 0.006 and 0.484

for the first and the second system, respectively. Meanwhile, the results generated by

the ANOVA test are also shown in Table 4.7, where the ANOVA test indicates a p-

value of 0.380 for the first system and 0.961 for the second system. In addition, the first

system (System IVa) produces the sensitivity of 0.8710, the specificity of 0.1111 and

the accuracy of 0.7000. The second system (System IVb) which implemented the Fuzzy

Histogram Equalisation generates better results where the sensitivity is 0.8387, the

specificity increased to 0.5556 and the accuracy is 0.7750. Based on the results

presented, it can be concluded that the implementation of the fuzzy preprocessing

techniques provides better contrast enhancement for fundus images, hence, it greatly

assists in the detection of microaneurysms, providing a more efficient and reliable

performance of the diagnosis system.

106

Figure 4.12 Snapshot of the user interface of System IV

Table 4.7 Summary results for Systems IV(a) and IV(b)

Features System IV(a) System IV(b)

Techniques Greyscale Conversion,

Contrast Limited

Adaptive Histogram

Equalisation, Circular

Hough Transform

Greyscale Conversion,

Fuzzy Histogram

Equalisation, Circular

Hough Transform

Number of fundus

images diagnosed

40 40

Confidence Interval 95% 95%

T-test p-value 0.006 0.484

T-test mean:

Expert

System

7.48

18.98

7.48

9.33

T-test standard

deviation:

Expert

System

10.268

21.021

10.268

11.425

T-test standard error

mean:

Expert

System

1.624

3.324

1.624

1.806

ANOVA test p-value 0.380 0.961

107

4.2.4 Automatic Detection of Microaneurysms Using Fuzzy

Histogram Equalisation, Fuzzy Filtering and Fuzzy Edge Detection

(System V)

Fuzzy image processing is a collection of different fuzzy approaches to the processing

of the images. Fuzzy edge detection, fuzzy histogram equalisation and fuzzy filtering

are among the fuzzy processing techniques that can be performed on images. In addition

to the fuzzy histogram equalisation, which was presented in Section 4.2.3, the fuzzy

filters for image filtering have been proposed by Patil and Chaudhari (2012), Toh

(2010) and also Kwan (2003). In addition to using fuzzy histogram equalisation and

fuzzy filters, fuzzy edge detection can also be performed.

In this section, in order to develop the proposed system, 600 fundus images from a

novel data set, collected from the Melaka Hospital, Malaysia (which is presented in

Section 3.9), have been used for evaluation. The system starts with the image

acquisition process, where the system selects images for further processing. The

selected images undergo preprocessing in order to improve image contrast in addition to

other enhancements. The preprocessed images are then used to locate and detect the

retinopathy signs, i.e., the microaneurysms. The numbers of microaneurysms detected

are displayed. Finally, in the classification phase, the images are classified into two

main cases: microaneurysms detected or microaneurysms not detected, based on the

number of microaneurysms detected. The overall process of microaneurysm detection is

presented in Figure 4.13.

108

Figure 4.13 Block diagram of System V for the automatic detection of

microaneurysms


The image preprocessing techniques involved in this present work include Greyscale

Conversion, Adaptive Histogram Equalisation, Fuzzy Histogram Equalisation, Fuzzy

Filtering and Fuzzy Edge Detection.

The first preprocessing technique used is the conversion of the colour fundus image into

a greyscale image, as greyscale is usually the best format for image processing.

Histogram equalisation is a computer image processing technique for improving the

image’s contrast. The proposed preprocessing techniques are used to help increase the

contrast of the colour fundus images. As mentioned in Section 4.2.3.1, for System IV,

the use of colour fundus images is more challenging compared to angiography and red-

free which are the other modes of fundus photography examination. Therefore,

appropriate techniques need to be implemented in order to improve the contrast of the

fundus images for better visualisation and detection. The Brightness Preserving

Dynamic Fuzzy Histogram Equalisation (BPDFHE) technique proposed by Sheet et al.

109

(2010), has been chosen as a preprocessing technique for the proposed detection of

microaneurysms in diabetic retinopathy screening. This is due to its good performance

and proven ability to work well for medical images, such as the pathology images

presented by Garud et al. (2011).

Image filtering is used to improve the image quality or to restore a digital image which

has been corrupted by some noise. The proposed system implemented the median filter

by employing the fuzzy techniques. The Noise Adaptive Fuzzy Switching Median

(NAFSM) filter is an extension to the Fuzzy Switching Median (FSM) filter developed

by Toh and colleagues (2010). It worked well both in removing salt-and-pepper noise

and preserving image details and textures, by incorporating fuzzy reasoning in

correcting the detected noisy pixel. The technique is composed of two modules, which

are the salt-and-pepper noise detection and the fuzzy noise cancellation module. This

technique is explained in Section 2.7, together with the fuzzy histogram equalisation

and fuzzy edge detection.

As outlined in Section 2.7, an edge is a boundary between two uniform regions. The

edge can be detected by comparing the intensity of the neighbouring pixels. However,

since uniform regions are not crisply defined, small differences of intensity between two

neighbouring pixels do not always represent an edge, it might instead represent a

shading effect. Therefore, the use of membership functions would overcome the

problems by defining the degree of pixel as to whether it belongs to an edge or to a

uniform region. The image gradients are created as the inputs of a Fuzzy Inference

System (FIS) for edge detection. For each input, a zero-mean Gaussian membership

function is specified, where if the gradient value for a pixel is 0 (region), then it belongs

to the zero membership function with a degree of 1. Another membership function is

added, which specifies the standard deviation from the zero membership function for

the gradient inputs. As an output of the FIS, which is the intensity of the edge-detected

image, the triangular membership functions (white and black) are specified. The values

of the standard deviation and also the values of the triangular membership functions for

the output can be changed in order to adjust the edge detector performance. Next, the

FIS rules can be specified to make a pixel white if it belongs to a uniform region,

110

otherwise the pixel presents as black. Figure 4.14 shows the membership functions of

the inputs and outputs for the edge detection purpose. For both inputs, which are the

gradient of every pixel on both axes, the parameters for the Gaussian membership

function consists of the central value or mean (0) and the standard deviation (0.1).

Meanwhile, for the outputs of the edge detection, which are the intensities of the edge-

detected image, there are three parameters required for the triangular membership

function. The parameters are the start, peak and end of the triangles of the membership

functions. The triplet parameter values for both white and black are (0.7, 0.9, 1) and (0,

1, 1), respectively. The fuzzy membership functions parameters were obtained by

changing the parameter values until better performance of the edge detector and

increased intensity of detected edges were produced. These parameters influence the

intensity of the detected edges. The Gaussian fuzzy membership function was used due

to the smoothness decision obtained, while the triangular function was chosen due to

simplicity and computational efficiency. The implementation of both fuzzy membership

functions for the edge detection provides better edge detection and more quality output

image. Figure 4.15 shows the output after each of the fuzzy pre-processing operations

on a selected image, as explained previously.

Figure 4.14 Membership functions for inputs and outputs

111

Figure 4.15 Preprocessing the output image with fuzzy approaches

4.2.4.2 Localisation and Detection of Microaneurysms

After performing the preprocessing techniques, both localisation and microaneurysm

detection takes place. In order to detect the circular microaneurysms satisfactorily, the

Circular Hough Transform (CHT) technique is implemented in the proposed system.

The implementation of the Circular Hough Transform uses the imdistline function in

Matlab to calculate the radius of microaneurysms in the fundus images. The viscircles

function that draws circles in the fundus image with a specified centre and radius onto

the current axes is presented in Section 4.2.2.2.


The counting of microaneurysm will then takes place. The presence of any number of

microaneurysms is represented as Diabetic Retinopathy, DR (microaneurysms

detected), and if there is no microaneurysm detected, it is considered as normal or no

retinopathy (microaneurysms not detected). The categorical types of the output whether

Normal/No DR or DR will be used later for the analysis of the system performance. The

categorical types from the expert diagnosis and the system generated are then compared.

The nominal inputs from both expert and system are represented as 1 for Normal/No

DR and 2 for DR. The average from the three experts will be used for the overall expert

diagnosis. Later, the comparison between the overall expert diagnosis and the proposed

system diagnosis will be completed and a statistical analysis of both performances will

then be produced.

(a) Fuzzy Histogram

Equalisation

(b) Fuzzy

Filtering

(c) Fuzzy Edge

Detection

112



performance analysis summary of the proposed system, which utilises different

techniques, is presented in Table 4.8. System V(a) implemented the greyscale

conversion, histogram equalisation and Circular Hough Transform (CHT), while

System V(b) proposed the implementation of the Fuzzy Histogram Equalisation in

addition to the greyscale conversion and CHT. System V(c) and System V(d) proposed

the implementation of other fuzzy image processing techniques, such as Fuzzy Filtering

and Fuzzy Edge Detection respectively. Two types of statistical tests were performed to

test the system performance compared to the expert diagnosis. Six hundreds fundus

images diagnosed by the expert and the systems are involved in both tests. The results

generated prove that, within the 95% confidence interval, the means of the annotated

images (1.54) and the system with the fuzzy techniques such as Fuzzy Histogram

Equalisation (1.49) and Fuzzy Edge Detection output (1.59) are not significantly

different. Thus, it can be concluded that the system findings are more likely to be no

different to the expert findings. However, the annotated images and the system with the

fuzzy median filter (with mean 1.43) are significantly different and show the opposite.

The inferential statistical analysis (i.e., T-test) result indicates a p-value of 0.00 (System

V(a)), 0.92 (System V(b)), 0.00 (System V(c)) and 0.73 (System V(d)), respectively.

The Chi-square test indicates a p-value of 0.00, 0.94, 0.00 and 0.81 for the four systems.

In addition, Table 4.8 shows that the results generated compare descriptively to the

methods of assessment between two groups: the expert and the system diagnosis for

both normal and diabetic retinopathy cases. The analysis based on the presented results

concludes therefore, that the implementation of the fuzzy preprocessing technique

provides better contrast enhancement and other improvements for fundus images, hence,

it greatly assists in detecting microaneurysms more efficiently therefore a reliable

performance of the diagnosis system can be produced.

113


Table 4.8 Summary results for System V(a), V(b), V(c) and V(d)

System V(a) System V(b) System V(c) System V(d)

Techniques Greyscale,

Histogram

Equalisation,

CHT

Greyscale, Fuzzy

Histogram

Equalisation,

CHT

Greyscale, Fuzzy

Median Filter,

CHT

Greyscale, Fuzzy

Edge Detection,

CHT

T-Test Mean:

Expert

System

1.54

1.67

1.54

1.49

1.54

1.43

1.54

1.59

T-Test p-value 0.00 0.92 0.00 0.73

Chi-Square p-value 0.00 0.94 0.00 0.81

Expert Count:

No DR

DR

276

324

276

324

276

324

276

324

System Count:

No DR

DR

201

399

305

295

343

257

246

354

114

4.3 Maculopathy and Diabetic Retinopathy Detection Maculopathy is represented by yellow lesions near the macula and is a disease of the

macula region of the retina. The macula is the centre of the retina and provides our

central vision. The macula region is a very sensitive area in the centre of which is the

fovea, a tiny area which is responsible for both detailed and colour vision (Taylor and

Batey, 2012). The detection of maculopathy therefore is very important because the loss

of vision at the fovea alone causes blindness. Maculopathy is present when there are any

exudates, haemorrhages or microaneurysms in the macula region. However, the visible

signs of maculopathy are only indirect markers for the possible presence of macula

oedema, which is the swelling of the retina (Taylor and Batey, 2012). The presence or

absence of the maculopathy will determine the need for treatment or referral. The

referral to the ophthalmologist is assigned if maculopathy is present. If maculopathy is

absent referral is not required and the screening will be repeated in one year’s time. The

combined detection of diabetic retinopathy and maculopathy therefore is vital in order

to effectively assist the management of diabetic retinopathy screening.

In order to identify maculopathy, the localisation and the detection of both macula and

fovea are essential, as maculopathy is represented by lesions in the macula region and

the fovea is at the centre of the macula. Kumar et al. (2013) proposed an approach in

detecting the macula by using bit plane decomposition and mathematical morphology

methods. Mubbashar et al. (2011) presented an automated system for the localisation

and detection of the macula in digital retinal images. The centring of the optic disc and

the blood vessel extraction are performed prior to the detection of the macula. This is

achieved by using the centre of the optic disc and by thresholding, followed by vessel

enhancement and then locating the macula as the darkest pixels in the region. In

addition, the detection of the macula is proposed by Akram et al. (2014) along with the

detection of exudates. The system proposed a contrast enhancement, thresholding and

blood vessel segmentation to detect the dark candidate, followed by the classification of

some of the set features for macula detection through the use of a Gaussian Mixtures

Model-based classifier. The localisation of the optic disc and the fovea in retinal fundus

images is also proposed by Sekhar et al. (2008). Morphological operations and the

Hough transform are implemented to locate the optic disc, while the fovea is located via

115

the spatial relationship with the optic disc and also from the spatial distribution of the

macula. They defined the region of interest as an area in a sector that originates at the

optic disc centre at an angle of 30° above and below the line between the optic disc

centre and the centre of the retinal image disc.

Meanwhile, Punnolil (2013) proposed an approach for the grading of diabetic

maculopathy through implementing the detection of retinal structures, such as the optic

disc, the macula and the fovea followed by the detection of lesions including the

exudates, haemorrhages and microaneurysms. These were later graded into classes of

diabetic maculopathy using a multiclass Support Vector Machines (SVM) classifier

based on the extracted features. Tariq et al. (2013) present a similar diabetic

maculopathy grading system starting with the optic disc localisation and the vascular

structure to extract the macula. Next, features such as area, compactness, mean

intensity, mean hue and others are extracted from the exudates detection and used to

classify different stages of maculopathy using a Gaussian Mixture Model-based

classifier. The detection of diabetic maculopathy in retinal images is also investigated

by Vimala and Kajamohideen (2014) through the use of morphological operations. The

detection of macula is generated from the image preprocessing techniques such as green

component extraction, filtering and adaptive histogram equalisation. Morphological

operations are then performed, such as top-hat transform and bottom-hat transform.

Other automatic gradings of diabetic maculopathy systems are proposed by

Siddalingaswamy and Prabu (2010), Hunter et al. (2011) and Chowriappa et al. (2013).

Siddalingaswamy and Prabu (2010) proposed the detection of the optic disc, fovea and

macula region, based on the location and diameter of the optic disc, followed by the

detection of the hard exudates using both clustering and mathematical morphological

techniques. The classification into severity levels of maculopathy, which consist of

mild, moderate and severe levels, is performed based on the location of exudates in the

marked macular region. An automated diagnosis of the maculopathy system is

presented by Hunter et al. (2011). Firstly, the optic nerve head and the fovea are

detected in order to locate the macula region. Candidate lesions are then segmented,

followed by feature extraction and, finally, classification by a multilayer perceptron. An

ensemble selection for the feature-based classification of diabetic maculopathy images

116

is suggested by Chowriappa et al. (2013). This is based on extracting textural features

and then using classifiers such as the hidden Naïve Bayes, Naïve Bayes, sequential

minimal optimisation (SMO) and the tree-based J48 algorithm to classify the features

into disease severity classes.

The detection of maculopathy is vital as it will eventually cause loss of vision if the

affected macula is untreated. Therefore, some researchers are focusing on this

challenging area and suggest several solutions for the detection of maculopathy on

retinal images. However, fuzzy processing has not been implemented during the

preprocessing stage within these previously reported maculopathy detection systems.

Therefore, the proposed system implements a combination of fuzzy techniques for the

image processing area of diabetic retinopathy and maculopathy detection. Two systems

have been proposed for detecting diabetic retinopathy alongside maculopathy in this

research work. The development of the first system (System VI) is explained in detail in

Rahim et al. (2015c), while the second system (System VII) is presented in Rahim et al.

(2016). Both systems have been evaluated with the 600 fundus images from the novel

data set, collected from the Melaka Hospital, Malaysia, which is presented in Section

3.9.

4.3.1 Detection of Diabetic Retinopathy and Maculopathy in Eye

Fundus Images Using Fuzzy Image Processing (System VI)

System VI implements a combination of fuzzy techniques in image preprocessing,

which involves fuzzy filtering, followed by fuzzy histogram equalisation and finally

fuzzy edge detection. System VI is different from System V presented in Section 4.2.4,

as System V proposed individual system variants implementing different fuzzy

processing techniques for the detection of microaneurysms. On the other hand, System

VI proposed a novel detection system for diabetic retinopathy and maculopathy by

combining several consecutive fuzzy image preprocessing techniques in one system,

based on previous encouraging results for the use of fuzzy image processing obtained

by System V. In addition, System VI is following the current practice observed by the

ophthalmologist in the classification and grading of diabetic retinopathy and

maculopathy, which classifies images into ten main classes, as explained in Section 3.9.

117

The system begins with the image acquisition process, where the system selects images

for further processing. The selected images undergo preprocessing in order to improve

the image contrast as well as perform other enhancements, with a combination of

several fuzzy techniques. The preprocessed images are then used for feature extraction,

where three features, namely, the area, the mean and the standard deviation of on pixels

are extracted. Finally, in the classification phase, some machine learning classifiers are

trained using these features to classify the images into their respective classes. Figure

4.17 presents the block diagram of the proposed system for the automatic screening and

classification of diabetic retinopathy and maculopathy using fuzzy image processing

techniques.

Figure 4.17 Block diagram of System VI for the automatic detection of diabetic

retinopathy and maculopathy using fuzzy image processing

118


Image preprocessing is the operation of improving the image data quality. Fuzzy

approaches are implemented in this proposed version of the system at the preprocessing

stage. The image preprocessing techniques involved in the present work include

Greyscale Conversion, Fuzzy Filtering, Fuzzy Histogram Equalisation and Fuzzy Edge

Detection. Similar proposed image preprocessing techniques, as in System V, are

implemented due to their good performance. In order to calculate the white pixels from

the edge-detected image, the output image is converted or inversed to produce the black

and white image. Figure 4.18 shows the membership functions of the inputs and outputs

for the edge detection purpose. For both inputs, which are the gradient of every pixel

on both axes, the parameters for the Gaussian membership function consists of the

central value or mean (0) and the standard deviation (0.1). Meanwhile, for the outputs of

the edge detection, which are the intensities of the edge-detected image, there are three

parameters required for the triangular membership function. The parameters are the

start, peak and end of the triangles of the membership functions. The triplet parameter

values for both white and black are (0, 1, 1) and (0, 0.7, 1), respectively. Figure 4.19

shows the output after each of the fuzzy pre-processing operations on a selected image.

Figure 4.18 Membership functions for inputs and outputs

119

Figure 4.19 Preprocessing of the output image

4.3.1.2 Feature Extraction

After performing the preprocessing techniques, feature extraction takes place to obtain

the features from the preprocessed images. Three preliminary features are proposed,

namely, the area of on pixels and the mean and standard deviation which will be

extracted for the purposes of detection. The first feature is the number of white pixels on

the black and white image, while the second and third features are the mean value and

the standard deviation of on pixels, respectively. These three shape-based features are

suitable to be extracted from the edge-detected inversed output image. For example, the

area of on pixels represents the detected maculopathy, which is the purpose of this

developed system. Other more sophisticated features could be extracted in addition to

these three proposed features in order to improve the classification performance.

(a) Original

image

(b) Greyscale

conversion

(c) Fuzzy

filtering

(d) Fuzzy

Histogram

Equalisation

(e) Fuzzy edge

detection

(f) Edge-detected

inversed

image

120


The extracted feature values were used in the classification stage, where the PRTools

Matlab toolbox (Duin et al., 2007) for pattern recognition was employed for this task. In

order to generate a variety of results and a good performance analysis of the system, two

types of classification were considered. First, the images have been classified into two

classes, i.e., normal (276 images) and diabetic retinopathy (324 images). In addition,

with the use of several machine learning classifiers, the images are then classified into

ten classes, which provide more details, i.e., No Diabetic Retinopathy (DR) with 276

images, and the other nine detailed classes of the DR cases, which are: Mild DR without

maculopathy (72 images), Mild DR with maculopathy (27 images), Moderate DR

without maculopathy (85 images), Moderate DR with maculopathy (83 images), Severe

DR without maculopathy (23 images), Severe DR with maculopathy (11 images),

Proliferative DR without maculopathy (6 images), Proliferative DR with maculopathy

(10 images) and Advanced Diabetic Eye Disease (ADED, with only 7 images). Several

machine learning classifiers, such as the binary decision tree classifier and the 1-nearest

neighbour classifier have been selected to train and classify images into these

categories.



performance of the proposed system, including the misclassification error, accuracy of

the individual classifiers and also the specificity and sensitivity of the two classifiers for

both categories are presented in Table 4.9, based on the confusion matrix generated.

Since the dataset is hugely imbalanced for some classes including mild DR with

maculopathy, severe DR with and without maculopathy, proliferative DR with and

without maculopathy and Advanced Diabetic Eye Disease (ADED), the minority classes

were oversampled. This was achieved by running the system with images from these

classes replicated several times in order to obtain various extracted feature values and

balance the dataset. The developed dataset is split randomly into 90% for training and

the remaining 10% for testing. The process is repeated ten times in a cross-validation

procedure in order to generate unbiased results. The average results on the ten runs for

each of the two classifiers are reported. The experimental results show that the two

121

classifiers, and especially the k-nearest neighbour, are able to identify well for both

main categories. The two classifiers identified the diabetic retinopathy cases much

better however in the two classes’ case, as there were more examples of such images in

the database compared to other example of the ten classes. The maculopathy can be

seen clearly from the inversed edge-detected image and the area of on pixels will have a

higher value for those images with maculopathy.


Table 4.9 Summary results for System VI

Category I : 2 classes Category II : 10 classes

Binary

decision tree

k-nearest

neighbour

Binary

decision tree

k-nearest

neighbour


Accuracy 0.7461 0.7861 0.5605 0.7025

Specificity 0.4536 0.5572 0.4500 0.6500

Sensitivity 0.8403 0.8598 0.5956 0.7297

122

4.3.2 Detection of Diabetic Retinopathy and Maculopathy in Eye

Fundus Images Using Fuzzy Image Processing and Retinal

Structures Segmentation (System VII)

System VII implements a combination of fuzzy techniques for image preprocessing,

such as fuzzy filtering and fuzzy histogram equalisation. In addition, the proposed

system implements the localisation and the detection of four retinal structures, which

are the optic disc, the blood vessels, the macula and the fovea, which are important in

the identification of maculopathy. Furthermore, the system detects diabetic retinopathy

lesions, the exudates, which are vital in the detection of exudative maculopathy. Several

features extracted from the exudates lesions and from the maculopathy are used for

classification. The system is evaluated with a combination of normal and diabetic

retinopathy images, including the maculopathy fundus images from the new dataset

developed, as discussed in Section 3.9.

The proposed diabetic retinopathy and maculopathy detection system has been created

using the Matlab R2014a environment. The system starts with the image acquisition

process, where the system selects images from the folder for further processing. Next,

the image preprocessing task takes place in order to improve the image quality. This

task includes the implementation of the fuzzy image processing techniques and the

retinal structures’ localisation and segmentation. In order to detect the maculopathy, the

lesions called exudates must be identified (where exudates situated in the macula region

represents the exudative maculopathy). Several features are then extracted from the

preprocessed image. Finally, in the classification phase, several machine learning

classifiers, such as the k-nearest neighbour, support vector machines and Naïve-Bayes

classifiers are trained using the generated features in order to classify the images into

their respective classes. Figure 4.21 shows the block diagram of the proposed system for

automatic screening and classification of diabetic retinopathy and maculopathy in

retinal images using fuzzy image processing techniques. The individual stages are

discussed in more detail in the following sections.

124

techniques: green channel extraction, fuzzy filtering and fuzzy histogram equalisation.

Figure 4.22 shows the output image obtained after the preprocessing stages.

The first preprocessing technique is the extraction of the green channel from the colour

fundus image, which consists of the red, green and blue channels. In the previous

proposed systems (System I – System VI), the colour fundus images are converted into

a greyscale image. In the present system, the green channel extraction is used to

investigate the difference and capability of the green channel format compared to both

the greyscale format and other image colour conversion formats. In most of the research

that has been carried out in diabetic retinopathy screening, the green channel is used for

detecting the diabetic retinopathy lesions such as haemorrhages, vessels and

microaneurysms. The red channel is somewhat saturated while the green channel

contains more structural information. Therefore, it is sensible to use the green channel

for the segmentation or morphological operations, since most of the image processing

tools work for greyscale only. Due to the capability of the green channel, the colour

fundus image is converted into the green plane channel. In addition, the aim of the

conversion is to investigate the ability of colour features, as one of the features to be

used for classification besides shape and intensity features. The next pre-processing

techniques are performed on the green channel image.

Image filtering needs to be implemented to improve the image quality or restore the

digital image which tends to have a variety of noise types. The poor photo quality may

also be due to equipment related factors, such as a dirty lens or a dirty computer screen.

In addition, distractions from the surroundings, such as too bright a room can be a factor

in generating poor quality fundus photographs. The filtering process therefore is

necessary to overcome this problem and to enable the image for further processing and

for effective grading task. The proposed system implements the median filter with fuzzy

techniques described by Toh et al. (2010) termed the Noise Adaptive Fuzzy Switching

Median (NAFSM) filter. Although the proposed technique is not working well as an

individual system variant in System V for microaneurysm detection, the technique has

been working well in System VI for both diabetic retinopathy and maculopathy

detection.

125

After filtering the image from any noise, the third preprocessing technique, the fuzzy

histogram equalisation, is performed on the images. The role of the histogram

equalisation is to improve the contrast of the image. As stated above, colour fundus

images are more challenging compared to angiography and red-free modes of fundus

photography examination. The implementation of the histogram equalisation by

employing fuzzy techniques therefore helps to improve the contrast of the fundus

images for better visualisation and detection. The technique called Brightness

Preserving Dynamic Fuzzy Histogram Equalisation (BPDFHE) proposed by Sheet et al.

(2010) was found to work well on colour fundus images (see System IV- System VI)

and has also been chosen as a preprocessing technique in this proposed system.


4.3.2.2 Retinal Structure Extraction

The extraction of retinal structures assists the diagnosis of eye diseases, through

improving both the detection results and grading for lesions. The detection of retinal

structures for diabetic retinopathy screening therefore, such as the optic disc, blood

vessels, the macula and the fovea are essential in order to produce a reliable screening

system. Figure 4.23 shows the output image with the extraction of retinal structures.

(a) Original image (b) Greyscale

conversion

(c) Green channel

extraction

(d) Green channel

complement (e) Fuzzy filtering

(f) Fuzzy histogram

equalisation

126

The optic disc is the most important retinal structure in the screening of diabetic

retinopathy. The optic disc has the largest high contrast among all the circular shape

areas. The location of the optic disc is of critical importance in retinal image analysis

and is required as a prerequisite stage in detecting lesions, including exudate detection.

Optic disc localisation and segmentation therefore are both performed in the proposed

system in order to improve the overall detection accuracy. The technique used in the

proposed system is based on the circular Hough transform (CHT), to ensure more

reliability in the optic disc detection task. In this case, the Hough transform for circles is

used to locate the optic disc, due to its circular shape. Based on a specified radius range,

the system finds circles in fundus images using CHT with the use of the imfindcircles

function in Matlab. After finding the circles in the image based on the radius range, the

function viscircles in Matlab is then used to create a circle. The function draws circles

with the specified centre and radius onto the current axes on the fundus images. The

CHT technique has also been used for the detection of microaneurysms in Systems III

and V, as this technique is works well in detecting circular shapes. In addition to the

proposed CHT technique, the polygonal Region of Interest (ROI) technique can be

performed on the fundus image to detect the optic disc. A region of interest is a portion

of an image for filtering purposes and also for performing other operation. In order to

implement this method, the roipoly Matlab function is used, which aims to create an

interactive polygon tool associated with the image displayed in the current image,

known as the target image. The function roipoly works by selecting vertices of the

polygon and, as a result, it returns a binary image that can be used for masked filtering.

The function allows moving, deleting or resizing the polygon as well as moving, adding

or deleting a vertex and also changing the colour of the polygon.

After locating the optic disc, another important retinal structure represented by the

blood vessels, is then extracted. There are several vessel extraction techniques

including: morphological operations, the Kirsch filter, the Frangi filter, local entropy

and entropic thresholding. For the proposed system, morphological operations are

implemented for the extraction of the blood vessels. The operation starts where the

fuzzy histogram equalisation output image is opened, using a disc shaped structuring

element. The image background is then removed and the image is thresholded to obtain

127

the binary images containing the vessels. Finally, the morphological open is

implemented using the bwareaopen function to remove any small noise. In order to

generate the elimination of the vessels, the binary image can be subtracted from the

Gaussian filtered image so that the final image has vessel free candidates.

The detection of the optic disc is vital for the detection of macula because its centre can

be used for macula detection. The macula appears as a dark region near the centre of the

image and its location is at a two-disc-diameter distance from the optic disc. The fovea

is the centre of the macula. Once the macula is identified, it is simple to determine the

fovea as its centre. The macular region can then be defined using two clinical

approaches. The first approach uses the Early Treatment Diabetic Retinopathy Study

(ETDRS), which indicates the presence of clinically significant macular edema (CSME)

as any part located within one disc diameter of the centre of the fovea (Wilkinson et al.,

2003). The second approach is based on the anatomical characteristics, where the

macular region is about 5-5.5 mm in diameter. In the proposed system, three methods

for macula detection are implemented. The first method uses the centre of the optic disc

identified by the circular Hough transform. The macula is identified based on the

distance specified from the optic disc. Once again, the circular Hough transform is used

to create a circle for the macula region. The detection of the macula region is proposed

by an angle of 37 degrees above and below the horizontal line crossing the optic disc

centre and with the centre of the macula region situated at 2 disc diameter (2DD)

distance from the optic disc centre. A circle of 1.3 discs diameter (1.3DD) is then

cropped to find lesions in the macula region. Figure 4.23(d) illustrates this process more

clearly. The cropped macula region can be used to locate larger and more appropriate

lesion areas relevant to maculopathy. The poly2mask function can be used to find the

region of interest of the macula region. In addition, the imcrop tool can also be used.

However, the cropping process has its limitations as it can be performed with

rectangular shapes only. This is because images are represented as arrays in Matlab and

arrays are required to be rectangular. The circular region of interest (ROI) cropping can

also be implemented with roicirclecrop function. The function crops the ROI in the

form of a circular shape and with a black background based on two points: the centre

and the radius of the circle.

129

4.3.2.3 Exudate and Maculopathy Detection

In order to detect maculopathy, exudates first need to be detected in the macula region.

The thresholding method is one of the ways to detect the exudate lesions. Thresholding

is used to extract an object from its background by assigning an intensity value

(threshold value) for each pixel, such as whether the pixel is classified either as an

object pixel or a background pixel. In this case, thresholding is used in the system

development as the simplest form of segmentation in order to segment the fundus image

into either an exudate or a background region. In the proposed system, the global

thresholding or fixed thresholding is implemented, as it is one of the more popular

methods. In global thresholding, the threshold value, T, is held constant throughout the

image. Choosing a correct threshold value is quite challenging and it is important to

ensure that the chosen value is neither too low nor too high. In order to find a suitable

threshold value for this system development, software termed ImageJ (ImageJ, 2014) is

used (through using the thresholding slider bar from the Adjust Threshold menu). From

the observations and the threshold value found during the searching process, it is

concluded that a value of 135 is the most appropriate in order to segment both the

exudate regions and the background of the fundus images, obtained after the fuzzy

histogram equalisation task. In addition to global thresholding, Fuzzy C-Means (FCM)

clustering can also be implemented in the system for detecting exudates. Morphological

operations can also be used for the extraction of exudates. The image after the optic disc

localisation and the segmentation of the blood vessels is then generated for the feature

extraction task, which will later be converted into a binary image in order to extract the

features of the exudates. After the detection of the exudates, the maculopathy can be

identified by cropping the area around the macula region. Figure 4.24(a) shows the

exudates extraction with the optic disc overlay, while Figure 4.24(b) shows the

maculopathy output within the proposed macula region.

130

Figure 4.24 Exudates and maculopathy extraction output image

4.3.2.4 Features Extraction

After performing the preprocessing techniques, the retinal structure extraction and also

the exudates detection, feature extraction then takes place in order to obtain the features

from the given images. There are three features selected from the exudates output image

and macula image, respectively. They are the area of on pixels, the mean of on pixels

and the standard deviation of the on pixels for the exudates region. In addition, there are

another three features from the maculopathy region, i.e., the area of on pixels, the mean

of on pixels and the standard deviation of the on pixels for the maculopathy. These

shape-values are chosen as they are suitable for representing the exudate and

maculopathy regions. Since the pre-processing techniques are performed after the

conversion of colour fundus images into the green channel, it can be concluded that the

system also investigated the capability of the colour-based features as well. The

difference between the mean pixel values for the exudate and maculopathy regions and

between the standard deviations of pixel values for the exudate and maculopathy

regions are explored. The features between the exudates region and the maculopathy

region are then compared. The value of the exudates area is higher compared to the area

of the maculopathy value as the maculopathy is located around the macula, and the

diameter of the macula is small.

(a) Exudates detection (b) Maculopathy detection

131


The six extracted features were used in the classification stage. As the system focuses

on the detection of maculopathy, the classes from the developed dataset, as presented in

Table 3.2 above (in Section 3.9), have been categorised into two classes, which are:

maculopathy detected or eye fundus with maculopathy (131 images) and maculopathy

not detected or fundus without maculopathy (469 images). Several machine learning

algorithms, such as the 1-nearest neighbour classifier, Naïve-Bayes classifier, support

vector machines (SVM) and radial basis function kernel SVM, have been selected to

train and classify images into their respective classes.


Figure 4.25 shows the user interface snapshot of the proposed system. The results are

shown in Table 4.10. The confusion matrix, the sensitivity, the specificity and the

accuracy of the individual classifiers are presented. Since the maculopathy detected

class (131 images) is imbalanced compared to the maculopathy not detected class (469

images), the maculopathy detected class was oversampled several times. As a result, a

total of 990 images, consisting of 469 images from the maculopathy not detected class

and 521 images from the maculopathy detected class are involved in the final

classification stage. The new dataset is split randomly into 90% for training and the

remaining 10% for testing purpose. The process is repeated ten times in a cross-

validation procedure to generate unbiased results. The results are averaged over ten runs

for each of the classifiers. The experimental results show that the four classifiers are

able to identify well for both categories, in particular the k-nearest neighbour and radial

basis function kernel support vector machine. The prior detection of exudates is

important because their presence will determine the whether or not maculopathy is

present. Based on the results presented, it can be seen that the sensitivity value is a little

lower than the specificity. This is due to the fact that the total number for maculopathy

detected (abnormal) cases is lower than the total number for maculopathy not detected

cases (normal), which was explained above. The other reason may be the capability of

the global thresholding technique which was used for the detection of the exudates. The

proposed threshold value may be quite high and, as a result, some of the exudates may

not be detected by the proposed technique.

132



k-nearest

neighbour

Polynomial

Kernel

SVM

RBF Kernel

SVM

Naïve-Bayes


Accuracy 0.9300 0.7000 0.9300 0.7500

Specificity 1.0000 0.9787 0.9362 0.9149

Sensitivity 0.8679 0.4528 0.9245 0.6038

4.4 Summary A preliminary system for the detection of diabetic retinopathy using a combination of

non-fuzzy techniques is presented in System I. Several individual systems for the

automatic detection of microaneurysms in colour fundus images for diabetic retinopathy

screening are also presented in Systems II to V. System II highlights the automatic

detection of microaneurysms in colour fundus images using segmentation and feature

extraction. Two subsystems for an automatic detection of microaneurysms in colour

133

fundus images using the Circular Hough Transform method for the localisation of

microaneurysms have been proposed, due to the capability of the proposed method to

detect circular shapes. The first subsystem proposed a combination of image processing

techniques and Circular Hough Transform, while the second subsystem presented the

detection of microaneurysms using fuzzy image processing. The first system, which

applies a non-fuzzy technique for image preprocessing is compared to the second

system, which implements a fuzzy image preprocessing technique. Based on the results

obtained, it was concluded that the implementation of fuzzy preprocessing techniques

provides better contrast enhancement for fundus images and in addition it can greatly

assist in detecting microaneurysms. Due to promising results in the implementation of

the fuzzy histogram equalisation technique for the detection of microaneurysms, the

development of the microaneurysms detection system has been enhanced by developing

a new dataset and using other fuzzy techniques for image preprocessing.

The implementation of separate fuzzy filtering and fuzzy edge detection, in addition to

the fuzzy histogram equalisation (as discussed previously), has been proposed for the

automatic detection of microaneurysms in System V. The fuzzy techniques work better

for both fuzzy histogram equalisation and fuzzy edge detection. The analysis shows that

the implementation of the fuzzy preprocessing techniques give better contrast

enhancement in addition to other improvements, such as brightness, and improved

segmentation for fundus images. The use of fuzzy image processing techniques plays an

important role in producing better image quality and improved performance analysis. In

addition, the capability of a combination of different fuzzy image processing techniques

for the detection of diabetic retinopathy and maculopathy in eye fundus images is

investigated in System VI. The proposed system implements a combination of fuzzy

techniques for image preprocessing, which combine fuzzy filtering, followed by the

fuzzy histogram equalisation and fuzzy edge detection. The system then classified

images into two classes including ten additional classes, which provide more detail

about the stages of the disease. The results show better identification of diabetic

retinopathy cases from the two-class-classifiers compared to the ten classes’ case. In

addition, the results show that maculopathy can be seen clearly from the generated

output image. Finally, System VII proposed a novel combination of fuzzy image

134

preprocessing techniques including the localisation of retinal structures (i.e., for the

optic disc, the blood vessels, the macula and the fovea), followed by feature extraction

and, finally, classification with some machine learning algorithms. To conclude, the use

of fuzzy image processing together with retinal structure extraction can help produce a

more reliable screening system for diabetic retinopathy.

Several methods and techniques of image processing were implemented to develop the

diabetic retinopathy screening system, which were presented and contrasted in the

sections above. Different techniques have been implemented for the proposed system in

order to either detect general retinopathy signs (i.e., whether retinopathy is present or

not), or more specific signs of retinopathy, such as microaneurysms, or maculopathy.

Although detection using colour fundus images was challenging, the proposed

techniques were able to improve the image contrast and eventually improve the

performance of the system. In addition, different public data sets have been used, which

contained a variety of colour fundus images. A good quality of available fundus images

can help in the process of the localisation and detection of the signs of retinopathy.

135

5 OVERALL RESULTS ANALYSIS AND DISCUSSION

This research has explored the development of the automatic systems for screening and

classification of diabetic retinopathy. Several systems have been successfully

developed, which contribute to achieve the objectives of this research. In order to

develop the system, eye fundus images are vital as an input. Therefore, sufficient good

quality images are required for both the development and the system performance. A

thorough system analysis and performance of the developed systems, alongside the

novel dataset analysis were presented.

This chapter presents the analysis of the results involving the data and the developed

systems. Section 5.1 discusses the analysis performed on the experts’ diagnosis as a

ground truth, including the descriptive and inferential analysis. The analysis of the

developed systems is presented in Section 5.2, which consists of two types of analysis,

namely the confusion matrix and statistical analysis. Meanwhile, Section 5.3 explains

some important points for further discussion based on the research findings. Finally,

Section 5.4 presents the summary of the chapter.

5.1 Developed Data Set Analysis As presented in Section 3.9, 600 fundus images from 300 patients were selected

together with the provided ground truth by the three experts from the Ophthalmology

Department, Melaka Hospital, Malaysia. The images were classified into ten

retinopathy stages, where the average from the three experts’ findings was used as the

final number of images for each stage. The first stage, which is normal (no retinopathy),

contributes the greatest number of 276 images. The abnormal or diabetic retinopathy

(DR) stage is divided into nine other stages: mild DR without maculopathy (72), mild

DR with maculopathy (27), moderate DR without maculopathy (85), moderate DR with

maculopathy (83), severe DR without maculopathy (23), severe DR with maculopathy

(11), proliferative DR without maculopathy (6), proliferative DR with maculopathy (10)

and finally, advanced diabetic eye disease, ADED (7).

136

The ground truth delivered by the experts can be placed into four categories. Table 5.1

shows the variety of the categorisation which can be used for system testing. The first

categorisation is the original classification made by the experts, which divides into ten

stages of retinopathy. The second categorisation divides into two cases, which are “no

diabetic retinopathy” and “diabetic retinopathy”. Meanwhile, the third categorisation

consists of four cases involving stages of no diabetic retinopathy, non-proliferative

diabetic retinopathy (mild, moderate and severe cases), proliferative diabetic

retinopathy and finally advanced diabetic eye disease. The fourth categorisation is based

on maculopathy detection, which classifies into two cases, namely maculopathy

detected and maculopathy not detected.

Table 5.1 Expert diagnosis summary categorisation

Categorisation I Categorisation II

Retinopathy Stage No. of

Images


Images

No DR 276 No DR 276

Mild DR without maculopathy 72 DR 324

Mild DR with maculopathy 27

Moderate DR without maculopathy 85

Moderate DR with maculopathy 83

Severe DR without maculopathy 23

Severe DR with maculopathy 11

PDR without maculopathy 6

PDR with maculopathy 10

ADED 7

Total 600 600

Categorisation III Categorisation IV


Images


Images

No DR 276 Maculopathy detected 131

Non-Proliferative DR 301 Maculopathy not detected 469

Proliferative DR 16

ADED 7

Total 600 600

Based on the ground truth provided by the experts, several descriptive and inferential

analysis tasks using the SPSS statistical package were performed.

137

i. Boxplot

The boxplot is the first method of assessment performed for our analysis. It is a useful

visualisation of how the data is distributed. In addition, the boxplot is able to display the

distribution of the variable scale and the pinpointing of the outliers. Moreover, the

boxplot shows five statistics, which are the minimum, the first quartile, the median, the

third quartile and the maximum value. Figure 5.1 shows the representation of the

boxplot performed on the ground truth, which involves a total of 1,800 images, i.e., 600

from each expert. It shows that for the first expert, there are several outliers, i.e., the

extreme values, which do not fall within the inner fences. The outliers mean that these

values are not frequent, particularly for the first expert, and that most of the time, these

values are not within the normal range. Meanwhile, the dark line shows the median,

which is the measure of the central tendency. It shows that for the first and third experts,

the median value is 1 (no diabetic retinopathy), while for the second expert the median

generated is 2 (mild diabetic retinopathy without maculopathy). The boxplot also shows

the 25th

percentile, representing 25% of cases or rows that have values below the 25th

percentile, in addition to the 50% of the cases or rows that lie both within the box and

the 75th

percentile.

Figure 5.1 Boxplot assessment

138

ii. Histogram

A histogram is an alternative for a descriptive analysis that can be performed for this

application. It is a visual summary of the distribution of values and it is useful for

showing the distribution of a single scale variable. Figure 5.2 shows the representation

of a histogram, where the 1,800 images, i.e., 600 images from each expert (600 images

from the first expert, 600 images from the second expert and another 600 images from

the third expert) are placed into the ten retinopathy stages. It shows that the majority of

the images are classified in the first stage, which is the “no retinopathy” stage, followed

by the fifth stage which is “moderate diabetic retinopathy with maculopathy”. The

histogram also generated the mean (2.83) and the standard deviation (2.269) for the

1,800 images used for the ground truth.

Figure 5.2 Histogram assessment

139

iii. Oneway ANOVA

The analysis of variance or the ANOVA test is an inferential analysis that can be

performed on the expert diagnosis ground truth. The test produces a one-way analysis of

variance for a quantitative dependent variable by a single factor or independent variable.

The analysis of variance is used to test the hypothesis that several means are equal. The

test generated shows a p-value of 0.000. The post-hoc test is then performed to

determine which means are different. Table 5.2 presents the output of the post-hoc test,

showing the multiple comparisons of the experts’ means. It can be concluded from

Table 5.2, that there are differences between the first expert and the second and third

expert, however there is also similarity between the second and the third expert.

Table 5.2 ANOVA multiple comparisons

iv. Chi-Square

The Chi-square test tabulates a variable into categories and computes a Chi-square

statistic. It compares the observed and expected frequencies in each category in order to

test whether or not all categories contain the same proportion of values. It can also test

that each category contains a user-specified proportion of values. In this case, the Chi-

square is useful for determining if there is a relationship among the experts. Table 5.3

shows the cross tabulation results of the Chi-square test. The inferential statistic with a

p-value of 0.000 indicates that there is no relationship among the three experts. Table

5.3 represents the number of images in each category for the three experts alongside the

percentage within the method of assessment (expert). The given percentage is calculated

by dividing the count for each of the retinopathy stages with the total numbers of

images for each expert – a total of 600 images. It can be seen that the distribution of

counts and percentages for each expert for no diabetic retinopathy stage is not

Method of Assessment p-value

Expert 1 Expert 2 0.001

Expert 3 0.001


Expert 3 1.000


Expert 2 1.000

140

significantly different. However, for mild diabetic retinopathy without maculopathy

stage for example, the count and percentage for the third expert (14, 2.3%) show a

significant difference compared to the first expert (55, 9.2%) and the second expert (58,

9.7%).

Table 5.3 Chi-square analysis

Results Method of Assessment

Expert 1 Expert 2 Expert 3

Count % Count % Count %

No DR 326 54.3 267 44.5 314 52.3

Mild DR without maculopathy 55 9.2 58 9.7 14 2.3

Mild DR with maculopathy 21 3.5 12 2.0 31 5.2

Moderate DR without maculopathy 79 13.2 90 15.0 43 7.2

Moderate DR with maculopathy 90 15.0 127 21.2 143 23.8

Severe DR without maculopathy 5 0.8 6 1.0 0 0.0

Severe DR with maculopathy 3 0.5 6 1.0 19 3.2

PDR without maculopathy 6 1.0 9 1.5 2 0.3

PDR with maculopathy 8 1.3 20 3.3 25 4.2

ADED 7 1.2 5 0.8 9 1.5

Total 600 100 600 100 600 100

p-value 0.000

5.2 System Results Analysis The results of the developed systems testing can be categorised into two types of

analysis. A confusion matrix (contingency table) is used to define the classification

model performance of test data. Meanwhile the statistical test is useful in providing a

mechanism for making quantitative decisions about a process or processes.

5.2.1 Confusion Matrix

The confusion matrix is chosen as a way to represent the results to analyse the

capability of the proposed system. The confusion matrix consists of information about

the actual and predicted classifications, namely, true positives, false negatives, false

positives and true negatives, which were explained earlier in Section 3.6 above. These

values will contribute to the calculation of the sensitivity (percentage of abnormal

fundus images classified as abnormal), specificity (percentage of normal fundus images

141

classifies as normal) and also the accuracy (overall correctness) of the proposed system.

The formulas of the sensitivity, specificity and accuracy calculation are shown in

Section 3.6. The confusion matrix is useful to present the classifier’s ability to correctly

predict or separate the classes and shows how the classification model predictions are

made. Table 5.4 shows the confusion matrix analysis for the developed systems. The

values of the true positives (TP), false negatives (FN), false positives (FP) and true

negatives (TN) from the generated confusion matrix for the first experiment from

System I (2 classes) and the first experiment from System VI (10 classes) are marked as

an example for how to calculate the sensitivity, specificity and the accuracy. The TP and

FN values are required for the calculation of the sensitivity, while the TN and FP values

are required for the specificity calculation.

5.2.2 Statistical Analysis

In addition to the confusion matrix, a statistical test can be performed to analyse the

system performance. Inferential statistical analysis, such as the analysis of variance

(ANOVA) test, t-test, Chi-square and other statistical tests can be performed. The

particular use of the statistical test depends on the nature of independent and dependent

variables analysed, such as categorical or continuous values, in addition to the number

of variables and normality. The different inferential statistical analysis presents the

variation of the system results analysis. Table 5.5 summarises the statistical tests output

on the developed systems results.

146

Overall, there are seven systems developed for the detection of diabetic retinopathy,

comprising one system for its general detection, four systems for the detection of

microaneurysms and finally, two systems for the detection of maculopathy. Figure 5.3

represents the workflow of all the developed systems, which provides better

visualisation and understanding about the developed systems proposed in this research

work. Firstly, a basic system for the detection of diabetic retinopathy using a

combination of non-fuzzy techniques is proposed. Next, four individual systems for the

automatic detection of microaneurysms in colour fundus images for diabetic retinopathy

screening are developed. The first system highlights the automatic detection of

microaneurysms in colour fundus images using segmentation and feature extraction.

Later, two subsystems for the automatic detection of microaneurysms in colour fundus

images using the Circular Hough Transform method for the localisation of

microaneurysms are proposed, due to the ability of the proposed method to detect

circular shapes. The first subsystem proposed a combination of image processing

techniques and Circular Hough Transform, while the second subsystem presented the

detection of microaneurysms using fuzzy image processing. The first subsystem, which

applies a non-fuzzy technique for image preprocessing, is compared to the second

system, which implements a fuzzy image preprocessing technique. Due to the promising

results in the implementation of the fuzzy histogram equalisation technique in the

detection of microaneurysms, the development of the microaneurysms detection system

is further enhanced by applying other fuzzy techniques in the image preprocessing of a

new dataset developed at Melaka Hospital. The implementation of fuzzy filtering and

fuzzy edge detection separately, in addition to the fuzzy histogram equalisation as

mentioned above, are proposed for the automatic detection of microaneurysms. The

capabilities of a combination of different fuzzy image processing techniques for the

detection of diabetic retinopathy and maculopathy in eye fundus images was

investigated (in System VI and System VII), where the proposed system implemented a

combination of fuzzy techniques in image preprocessing, which combine fuzzy filtering,

followed by the fuzzy histogram equalisation and fuzzy edge detection. The final

system proposed a combination of fuzzy techniques for image preprocessing, in

addition to the localisation and detection of four retinal structures, which are the optic

disc, the blood vessels, the macula and the fovea.

148

5.3 Discussion

The findings of this study, pertaining to the development of the effective diabetic

retinopathy automatic screening and classification using eye fundus images show, some

common issues:

i. Imbalanced dataset

A balanced dataset is an advantage for the purpose of classification. In the field of

medical imaging however, it is difficult to achieve such a balanced dataset as there are

not usually a large number of normal cases. This limitation therefore presents problems

in the classification phase. For diabetic retinopathy diagnosis, this problem of imbalance

also occurs in the available public databases, as stated in Section 3.8. For example, the

popular dataset Standard Diabetic Retinopathy Database Calibration Level 0

(DIARETDB0), offers only 20 normal images compared to 110 images with signs of

diabetic retinopathy. Meanwhile, the Standard Diabetic Retinopathy Database

Calibration Level 1 (DIARETDB1) consists of only five normal images while the

remaining 84 images are diabetic retinopathy case. In order to overcome this limitation,

duplication is required, where the minority class is oversampled several times in order

to avoid the imbalance problem. The ensemble of classifiers is another solution that can

overcome the imbalanced dataset limitation.

As presented in this research work, a new dataset was introduced. The new dataset

managed to overcome this imbalanced data problem, in that it provides 276 normal

images and 324 diabetic retinopathy images; almost a balanced total number of No

DR/Normal and DR/Abnormal images. However, the imbalance problem still occurs

when dealing with the more detailed categorisation, which classifies data into ten cases

(no retinopathy, mild DR without maculopathy, mild DR with maculopathy, moderate

DR without maculopathy, moderate DR with maculopathy, severe DR without

maculopathy, severe DR with maculopathy, proliferative DR without maculopathy,

proliferative DR with maculopathy and advanced diabetic eye disease). The total

numbers of severe and proliferative cases are significantly imbalanced due to the fact

that in severe cases, the patient will be referred directly for laser treatment rather than

149

for the capturing of their fundus images. Another reason is that diabetic retinopathy

takes a long time to progress to the severe stage of diabetes mellitus.

ii. Detection of exudates

As presented in Section 4.3 the localisation of exudates are important aspects to be

considered for detecting maculopathy. The detection of maculopathy is based on the

exudates finding, due to the fact that the maculopathy represents the exudates found in

the macula region. Therefore, the detection of exudates, as an input, is vital to ensure

the efficient detection of maculopathy. In addition, the fovea and the region of macula

detection is another important aspect that should be accounted for.

The research work implemented a thresholding method to detect the exudate lesions in

System VII. Since maculopathy detection depends on the successful location of

exudates, a precise exudate detection method is required. The proposed threshold values

used for the global thresholding should be both suitable and applicable to all the tested

images, as each image consists of different contrast values. An inappropriate threshold

value will result in a poor detection rate for exudates.

iii. Fuzzy techniques combination capability

Several fuzzy techniques have been implemented in the research work used for the

automatic screening and classification of diabetic retinopathy. Among the proposed

fuzzy techniques are the fuzzy histogram equalisation, fuzzy filtering and fuzzy edge

detection. Fuzzy median filtering was proved not to work well individually in the

detection of microaneurysms in System V(c). Fuzzy histogram and fuzzy edge detection

however were proved to work well individually in System V(b) and System V(d),

respectively.

On the other hand, the fuzzy median filter was working well with the combination of

other fuzzy techniques, such as the fuzzy histogram equalisation and fuzzy edge

detection. These combinations of fuzzy techniques were working well for the detection

of diabetic retinopathy and maculopathy as presented in System VI. Other fuzzy

filtering techniques should also be proposed and investigated in order to overcome the

150

limitations of the Noise Adaptive Fuzzy Switching Median (NAFSM) filter, which was

used in the proposed system.

iv. Variety of diabetic retinopathy categorisation

It can be seen that for the diabetic retinopathy and maculopathy detection, there are two

types of classification involved. The first categorisation is based on two cases, which

are termed “normal” and “diabetic retinopathy”. The second categorisation is more

detailed compared to the first as it consists of ten cases, i.e., No Diabetic Retinopathy

(DR) and the other nine detailed cases of diabetic retinopathy: Mild DR without

maculopathy, Mild DR with maculopathy, Moderate DR without maculopathy,

Moderate DR with maculopathy, Severe DR without maculopathy, Severe DR with

maculopathy, Proliferative DR without maculopathy, Proliferative DR with

maculopathy and Advanced Diabetic Eye Disease (ADED). These two types of

categorisation offer a variety of results and performance analysis of the system.

The most important reason for implementing the second categorisation (the ten cases) is

in order to follow the current practice observed by ophthalmologists in the diabetic

retinopathy and maculopathy classification and grading. This detailed categorisation has

been acknowledged by the experts during the data collection and the ground truth

preparation process.

v. Variety of system configurations

A variety of system configurations have been presented in this research work, which

aim to develop an automatic screening and classification system for diabetic

retinopathy, in order to reduce the workload for the screening team. It can be concluded

therefore that all the proposed systems are related. The research work began with the

preliminary system for automatic screening and the classification of diabetic retinopathy

using fundus images. The research work continued with the detection of one of the

diabetic retinopathy important features, which are microaneurysms. Since the detection

of maculopathy is vital for the urgency of referral, the research work has continued with

the detection of diabetic retinopathy and maculopathy based on current practice for

151

ophthalmologists. The earlier proposed techniques are reused for the following system

configurations due to their good performance.

Other researchers focus on the maculopathy detection on the eye fundus images and

suggest several solutions. They have proposed a variety of ways to detect and classify

the fundus images into different stages of maculopathy, such as mild, moderate and

severe maculopathy. However, this research work presented different mechanisms of

maculopathy detection based on ophthalmologists’ practice, which combine the

detection of both diabetic retinopathy and maculopathy, based on the discovery of

diabetic retinopathy signs. The severity level is based on the diabetic retinopathy

features, rather than the severity of the maculopathy. The classification refers to whether

maculopathy is present (with maculopathy) or not (without maculopathy). As a result,

the generated cases are: No Diabetic Retinopathy (DR), Mild DR with/without

maculopathy, Moderate DR with/without maculopathy, Severe DR with/without

maculopathy, Proliferative DR with/without maculopathy and Advanced Diabetic Eye

Disease (ADED). This categorisation is useful as two important aspects can be detected

in just one screening process. Moreover, the urgency of the referral which should

happen within four weeks, as presented in Table 2.14 by the National Institute for

Clinical Excellence, is applied to those who have any form of maculopathy, whether

mild, moderate or severe. The severity of the maculopathy is therefore not significant in

this case, provided that its presence or absence has been determined.

5.4 Summary

To conclude, this chapter has discussed the analysis performed on the developed data.

The overall analyses of the developed systems are also presented. Various techniques

can be used to analyse the data. The confusion matrix is one of the ways to represent the

performance of the classification task. It would also be useful to determine the

performance of the screening system. In addition, statistical analysis is an alternative

way to represent the relationship and capability of the data and the proposed system

results. The chapter also presented some common issues resulting from the findings of

the research.

152

6 CONCLUSIONS AND FUTURE WORK

The development of an automatic diabetic retinopathy screening system is a very

challenging task. The system development involves not only an advanced understanding

of image processing procedures, but also requires essential medical input, including

expert knowledge related to diabetic retinopathy and its screening procedure, in addition

to the eye fundus photography process. Efficient and cost-effective approaches in the

field of digital retinal imaging should be established. Diabetic retinopathy and

maculopathy screening is necessary to identify persons at risk of visual impairment.

Effective screening of diabetic retinopathy therefore is vital for early action, alongside

an effective diabetic complications preventive management.

This chapter concludes the thesis and provides a summary of the research. Section 6.1

reviews the performance of the research work, discusses the research contribution and

further reiterates the advancement offered by this thesis. Section 6.2 presents

suggestions for future research work that can provide further improvements in the field,

in addition to directions for such future research studies. Finally, Section 6.3

summarises and concludes this chapter.

6.1 Review of the Work and Contributions

The automatic grading of diabetic retinopathy is a rising research field aimed to

decrease the workload inherent in the conventional grading process. An automatic

diabetic retinopathy system would enable faster and more efficient diagnosis. Moreover,

preventive action to protect vision can be taken earlier. The output results generated

from the research work have answered the primary research aim detailed in Chapter 1,

where the computer-based imaging tool developed through this research has proven to

be effective in detecting important diagnostic features and classifying individuals into

the correct retinopathy stage.

The research main objective was to develop an automatic system for the diabetic

retinopathy detection and to classify images into retinopathy stages, based on diabetic

153

retinopathy severity standard. In order to achieve this, the research can be summarised

as follows:

i. The development of an automatic diabetic retinopathy screening system for

general detection

ii. The development of an automatic microaneurysms detection system, the first

important signs of diabetic retinopathy.

iii. The development of an automatic diabetic retinopathy and diabetic

maculopathy detection systems.

The introductory chapter of the thesis highlighted the issues of diabetic retinopathy

screening globally and in Malaysia in particular. Based on the objectives previously

presented, the following represent the undertakings and the main contributions of this

research:

The research developed several systems for the automatic diabetic

retinopathy detection, in addition to the identification of features of diabetic

retinopathy and maculopathy lesions (Rahim et al., 2014; 2015a; 2015b;

2015c; 2016)

The developments of the system introduced novel use of image processing

techniques for the fundus images handling in the screening process of

diabetic retinopathy (Rahim et al., 2014; 2015a; 2015b; 2015c; 2016)

A new dataset of fundus images was developed, as an input for the research

development, which is not only beneficial in the screening of diabetic

retinopathy process, but for other retinal disease screenings, such as

cataracts, glaucoma and hypertensive retinopathy among others. The new

dataset is available at http://creative.coventry.ac.uk/fundus (Rahim et al.,

2015b; 2015c; 2016)

A thorough system performance analysis was undertaken, which compared

the performance of automatic systems to the manual diagnosis performed by

the expert

154

Based on the proposed techniques and system configurations, the research

designed a complete process for the automatic diabetic retinopathy

screening, where the proposed system development framework can also be

implemented for the automatic detection of other retinal diseases with only

slightly different features involved

The research developed several automatic detection systems for the

microaneurysms, the earliest visible and most important sign of diabetic

retinopathy (Rahim et al., 2015a; 2015b)

The development of the systems employed fuzzy image processing

techniques as a core contribution to the performance of the research work

(Rahim et al., 2015a; 2015b; 2015c; 2016)

Results from this research will benefit a number of areas; most notably, it will

significantly improve the detectability of human eye diabetic retinopathy. Preventative

action therefore can be taken to decrease the rates of diabetic retinopathy problems, in

addition to the blindness risk. It also aims to assist clinicians to diagnose diabetic

retinopathy at an early stage by using the developed detection techniques. The decision

support system for clinical diagnosis would contribute greatly in assisting with the

management and detection of diabetic retinopathy. The automated system will assist an

ophthalmologist (or optometrist) to detect diabetic retinopathy (and its detailed

classification) in a more efficient and faster way, compared to manual analysis. In

addition to the classification of diabetic retinopathy, the system will indirectly assist in

the provision of recommended follow-up schedules and treatment for each category of

diabetic retinopathy, based on the generated classification. The crucial aim of diabetic

retinopathy is to identify any sight threatening conditions and to make sure prompt

treatment in order to avoid vision loss. Such appropriate ophthalmologist referrals

should therefore be facilitated with the adoption of this system.

Moreover, this research provides a significant contribution to knowledge in the area of

medical image processing. This is achieved by proposing a novel use of image

processing techniques in diabetic retinopathy screening, as listed in Chapter 1 and above

in this chapter. The research proposes the image processing techniques combination for

155

the general diabetic retinopathy detection (System I). In addition, the research proposed

a fuzzy technique-based eye screening system to detect the microaneurysms (System II,

System III, System IV and System VI), one of the most important diabetic retinopathy

signs. The developed system for the detection of microaneurysms could be a benchmark

for the detection of other diabetic retinopathy signs development systems, such as

exudates, haemorrhages and neovascularisation. Furthermore, the research presents a

novel automatic diabetic retinopathy and maculopathy detection in retinal images using

fuzzy image processing (System VI and System VII), which classifies retinopathy into

stages based on the actual practice of the screening team (i.e., drawing from

conventional practices). The proposed fuzzy-based image processing decision support

system will help the screening of diabetic retinopathy and decrease the workload of the

screening personnel.

The development of an online novel dataset, which consists of 600 colour fundus

images, alongside expert diagnosis is another important contribution of the research

(System V, System VI and System VII). The normal and diabetic retinopathy fundus

images combination from a new dataset representing the South East Asian population,

in particular the Malaysian demographic, was a major part of this research work. One

significant attribute of the dataset is that unlike other publicly available datasets, it

represents the Asian population, as large numbers of such data are uncommon. The

developed dataset is useful for researchers in promoting research on the area of retinal-

imaging, particularly in the field of diabetic retinopathy screening.

6.2 Future Work As highlighted in an earlier section, the research work has produced several novelties

that were embedded in the variants of the developed systems. Despite the fact that the

promising results in the several techniques implementation showed in the this study, it

is important to state that there is still room for improvement and some new directions

for future research are listed below.

156

i. Combination of diabetic retinopathy features detection

Diabetic retinopathy screening is a complicated process, as there are various signs of

that need to be identified and detected in order to achieve a complete screening system.

This research study however, concentrated on microaneurysms, the earliest and the most

important diabetic retinopathy signs. Other important diabetic retinopathy features, such

as exudates, haemorrhages, cotton wool spots and neovascularisation also need to be

identified, according to the standard clinical severity scales for diabetic retinopathy. For

each of the diabetic retinopathy features listed above, novel use of techniques should be

proposed for their effective detection. In addition to these features, there are some

exclusion criteria that need to be considered for the task of diabetic retinopathy

detection, for example polyps, bleeding and drusen. The combination of all these

features of diabetic retinopathy into a single system will provide a complete automatic

diabetic retinopathy detection system. The diabetic retinopathy screening process has

immense scope because it involves many features and criteria that need to be taken into

account. The research can be extended to identify the other remaining diabetic

retinopathy features and combine them into a complete screening system. A complete

and accurate system could be used to assist the team of diabetic retinopathy screening to

perform in an improved and more efficient way. As a conclusion, diabetic retinopathy

screening is not an easy task, as many criteria need to be thoroughly examined before

any diagnoses can be made.

ii. Counting constraints for microaneurysms

The counting of microaneurysms by an expert is required for a precise analysis in the

development of the system or model, in order to make a comparison between the system

results and the system counting. The number of microaneurysms given by the expert for

the novel developed dataset is required as an input for the statistical tests, for example

the T-test and Chi-square test. However, according to the expert from the Department of

Ophthalmology, Melaka Hospital, Malaysia, the microaneurysms counting is almost

impossible due to the following constraints. For example, if microaneurysms are

present, they are numerous in number and the counting would no longer be accurate.

This is further complicated when there is some overlapping among the microaneurysms.

Furthermore, another constraint is the fact that microaneurysms could be easily

157

confused with blot haemorrhages of about the same size. A solution would be to group

the microaneurysms into two main classifications, which are “No” for microaneurysms

not detected and “Yes” for microaneurysms detected. For the “Yes” choice, would be

another three sub-choices, i.e., “Low”, “Moderate/Medium” and “Severe”, according to

the approximate microaneurysms number estimated by the expert. If there are just a few

microaneurysms detected, then it is classified as “Low”, if there are several, it is

classified as “Moderate”, and if there are many, then the classification will be “Severe”.

The choice of these three cases is dependent on the threshold value for each class which

can be fixed in advance. Based on the three sub-choices of microaneurysms detected,

fuzzy logic may be used. This approach is an alternative to overcome the constraints to

the accurate counting of microaneurysms. It can be applied to a novel developed

dataset, that can in turn be proposed for future use of an automatic microaneurysms

detection system development.

iii. Combination of other fuzzy image processing techniques

In addition to the proposed techniques of image processing, the research can be

improved by the implementation of other different preprocessing techniques

combinations, including those based on fuzzy approaches. The fuzzy image processing

can be applied to fully explore the variety of fuzzy techniques capabilities. In order to

produce a more reliable screening system, fuzzy image processing in addition to the

extraction of retinal structure can be employed in diabetic retinopathy screening. An

alternative technique for the future detection of optic disc, Fuzzy Circular Hough

Transform, will be implemented.

iv. Widely accessible developed dataset

The research is not merely proposing an automatic diabetic retinopathy screening

detection system, but has also introduced a new dataset of eye fundus images. The new

dataset would be beneficial to researchers and practitioners working in the retinal

imaging area, especially the diabetic retinopathy screening field. For future work, the

online dataset can be made more widely accessible by integrating the dataset with other

popular databases, for example the University of California Irvine (UCI) Machine

Learning Repository, among others.

158

6.3 Summary One of the main health threats is diabetes mellitus, as it leads to a severe and enduring

complications, as well as sight-threatening conditions. Diabetic retinopathy is a diabetes

complication that affects the blood vessels damage inside the retina. Thus, both initial

detection and timely treatment is essential for this retinal problem. Such an effective

diagnosis and the diabetic retinopathy grading can assist in early detection of diabetic

retinopathy and may decrease its prevalence. It is envisaged that a decision support

system for clinical screening would contribute to and greatly assist in the management

as well as the detection of diabetic retinopathy. It is also hoped that the developed

detection technique will assist clinicians to diagnose diabetic retinopathy at an early

stage.

This research project examined the use of fundus images and techniques of image

processing to detect the diabetic retinopathy presence in the eye. This is a particularly

challenging problem and this research has made novel use of image processing

techniques to automatically detect the retinopathy stages. Highly efficient and accurate

image processing techniques must be used in order to produce an effective diagnosis of

diabetic retinopathy. As such, this research proposed a new mechanism that could

provide ophthalmologists with a novel way to identify and treat those who are most

vulnerable to vision loss from diabetic retinopathy. Specific image processing

techniques have been proposed and developed to test the efficiency of this approach

compared to others. As a conclusion, the implementation of fuzzy image processing

techniques play a significant role in generating better quality of image and enhanced

performance. Eventually it can contribute to producing a more reliable screening

system.

REFERENCES

Abdelazeem, S. (2002) ‘Microaneurysm Detection Using Vessels Removal and Circular Hough

Transform’. in Proceedings of the Nineteenth National Radio Science Conference (NRSC 2002).

held 19-21 March 2002 at Alexandria. USA: IEEE, 421-426

Acharya, U. R., Dua, S., Du, X., Sree, V. S., and Chua, C, K. (2011) ‘Automated Diagnosis of Glaucoma

Using Texture and Higher Order Spectra Features’. IEEE Transactions on Information

Technology in Biomedicine 15 (3), 449-455

Adal, K. M., Ali, S., Sidibe, D., Karnowski, T., Chaum, E., and Meriaudeau, F. (2013) ‘Automated

Detection of Microaneurysms Using Robust Blob Descriptors’. SPIE Medical Imaging-

Computer Aided Diagnosis Feb 2013, 8670-22

Adal, K. M., Sidibe, D., Ali, S., Chaum, E., Karnowski, T, P., and Meriaudeau, F. (2014) ‘Automated

Detection of Microaneurysms Using Scale-adapted Blob Analysis and Semi-supervised

Learning’. Computer Methods and Programs in Biomedicine 114, 1-10

Addoor, K. R., Bhandary, S. V., Khanna, R., Rao, L. G., Lingam, K. D., Binu, V. S., Shivaji, S., and

Nandannaver, M. (2011) ‘Assessment of Awareness of Diabetic Retinopathy Among the

Diabetics Attending the Peripheral Diabetic Clinics in Melaka, Malaysia’. Medical Journal

Malaysia 66 (1), 48-52

Ahmad, F. M. H., Izhar, L. I., Venkatachalam, P. A., and Karunakar, T. V. (2007) ‘Extraction and

Reconstruction of Retinal Vasculature’. Journal of Medical Engineering & Technology 31 (6),

435-442

Akram, M. U., Tariq, A., Khan S. A., and Javed, M. Y. (2014) ‘Automated Detection of Exudates and

Macula for Grading of Diabetic Macular Edema’. Computer Methods and Programs in

Biomedicine 114, 141-152

Akram, M. U., Jamal, I., Tariq, A., and Imtiaz, J. (2012) ‘Automated Segmentation of Blood Vessels for

Detection of Proliferative Diabetic Retinopathy’. in Proceedings of IEEE-EMBS International

Conference on Biomedical and Health Informatics (BHI 2012), ‘Global Grand Challenge of

Health Informatics’. held 2-7 January 2012 at Hong Kong and Shenzhen, China. USA: IEEE,

232-235

Akram, M. U., Khalid, S., Tariq, A., Khan, S. A., and Azam, F. (2014) ‘Detection and Classification of

Retinal Lesions for Grading of Diabetic Retinopathy’. Computers in Biology and Medicine 45,

161-171

Akram, M. U., Khalid, S., and Khan, S. A. (2012) ‘Identification and Classification of Microaneurysms

for Early Detection of Diabetic Retinopathy’. Pattern Recognition 46, 107-116

Akram, M. U., Khalid, S., Tariq, A., and Javed, M. J. (2013) ‘Detection of Neovascularization in Retinal

Images Using Multivariate m-Mediods Based Classifier’. Computerized Medical Imaging and

Graphics 37, 346-357

This

item

has bee

n removed

due to

3rd Party Copyright. The unabridged version of the thesis can be found in

the Lancester Library, Coventr

y University.

160

Al-Diri, B., Hunter, A., Steel, D., Habib, M., Hudaib, T., and Berry, S. (2008) ‘REVIEW- A

Reference Data Set for Retinal Vessel Profiles’. in Proceedings of the 30th Annual International

Conference of the IEEE Engineering in Medicine and Biology Society, ‘Personalized Healthcare

through Technology’. held 20-24 August 2008 at Vancouver, Canada. USA: IEEE, 2262-2265

Alipour, S. H. M., Rabbani, H., and Akhlaghi, M. R. (2012) ‘Diabetic Retinopathy Grading by Digital

Curvelet Transform’. Computational and Mathematical in Medicine, 1-11

Al-Rawi, M., Qutaishat, M., and Arrar, M. (2007) ‘An Improved Matched Filter for Blood Vessel

Detection of Digital Retinal Images’. Computers in Biology and Medicine 37, 262-267

American Academy of Ophthalmology Retina Panel (2008) Preferred Practice Pattern Guidelines.

Diabetic Retinopathy. San Francisco: American Academy of Ophthalmology

Antal, B., and Hajdu, A. (2010) ‘Improving Microaneurysm Detection in Color Fundus Images by an

Optimal Combination of Preprocessing Methods and Candidate Extractions’. in Kleijn, B. (ed.)

Proceedings of the 18th European Signal Processing Conference EUSIPCO-2010’. held 23-27

August 2010 at Aalborg, Denmark. Greece: European Association for Signal, Speech and Image

Processing (EURASIP), 1224-1228

Antal, B., and Hajdu, A. (2013) ‘Improving Microaneurysm Detection in Color Fundus Images by Using

Context-aware Approaches’. Computerized Medical Imaging and Graphics 37, 403-408

Aravind, C., Ponnibala, M., and Vijayachitra, S, (2013) ‘Automatic Detection of Microaneurysms and

Classification of Diabetic Retinopathy Images Using SVM Technique’. IJCA Proceedings on

International Conference on Innovations in Intelligent Instrumentation, Optimization and

Electrical Sciences ICIIIOES (11), 18-22

Awang, Z. (2012) Research Methodology and Data Analysis. 2nd edn. Selangor: UiTM Press

Bala, M. P., and Vijayachitra, S. (2012) ‘Computerised Retinal Image Analysis to Detect and Quantify

Exudates Associated with Diabetic Retinopathy.’ International Journal of Computer

Applications 54 (2), 7-12

Bhattacherjee, A. (2012) Social Science Research: Principles, Methods and Practices. Florida: Global

Text Project

Burges, C. J. C. (1998) ‘A Tutorial on Support Vector Machines for Pattern Recognition’. Data Mining

and Knowledge Discovery 2 (2), 121-167

Centre for Eye Research Australia (2013) Diabetic Retinopathy [online] available from

<http://www.cera.org.au/uploads/pdf/FS_DR.pdf> [4 July 2013]

Chowriappa, P., Dua, S., Rajendra, A. U., and Muthu, R. K. M. (2013) ‘Ensemble Selection for Feature-

based Classification of Diabetic Maculopathy Images’. Computers in biology and medicine 2013

43 (12), 2156-2162

Creswell, J. W. (2009) Research Design Qualitative, Quantitative and Mixed Methods Approaches. 3rd

edn. United States of America: SAGE Publications, Inc

Creswell, J. W. (2012) Educational Research Planning, Conducting and Evaluating Quantitative and

Qualitative Research. 4th edn. United States of America: Pearson Education, Inc

161

Duin, R. P. W., Juszczak, P., Paclik, P., Pekalska, E., de Ridder, D., Tax, D. M. J., and Verzakov, S.

(2007) ‘PRTools4.1, A Matlab Toolbox for Pattern Recognition’. Delft University of

Technology

Early Treatment Diabetic Retinopathy Study Research Group (1991) ‘Grading Diabetic Retinopathy

From Stereoscopic Color Fundus Photographs-An Extension of the Modified Airlie House

Classification’. ETDRS report number 10. Ophthalmology 98 (5 Suppl), 786-806

Er, M. J., Wu, S., and Gao, Y. (2003) Dynamic Fuzzy Neural Networks Architecture, Algorithms and

Applications. Singapore: Mc Graw Hill Education

Esmaeili, M., Rabbani, H., Dehnavi, A. M., and Dehgani, A. (2012) ‘Automatic Detection of Exudates

And Optic Disk in Retinal Images Using Curvelet Transform’. IET Image Processing 6 (7),

1005-1013

eyeSmart (2014) What Is Diabetic Retinopathy [online] available from <http://www.geteyesmart.org/

eyesmart/diseases/diabetic-retinopathy/> [10 July 2014]

Fleming, A. D., Philip, S., Goatman, K. A., Olson, J. A., and Sharp, P. F. (2006) ‘Automated

Microaneurysm Detection Using Local Contrast Normalization and Local Vessel Detection’.

IEEE Transactions on Medical Imaging 25 (9), 1223-1232

Gardner, G. G., Keating, D., Williamson, T. H., and Elliot, A. T. (1996) ‘Automatic Detection of Diabetic

Retinopathy Using an Artificial Neural Network: A Screening Tool’. British Journal of

Ophthalmology 80, 940-944

Garud, H., Sheet, D., Suveer, A., Karri, P. K., Ray, A. K., Mahadevappa, M., and Chatterjee, J. (2011)

‘Brightness Preserving Contrast Enhancement in Digital Pathology’. in Siddavatan, R., and

Ghrera, S. P. (ed.) Proceedings of the 2011 International Conference on Image Information

Processing (ICIIP 2011). held 3-5 November 2011 at Himachal Pradesh, India. USA: IEEE, 1-5

Goh, P. P., Elias, H., Norfariza, N., and Mariam, I. (2008) ‘National Eye Database- A Web Based

Surveillance System’. Medical Journal Malaysia (63) Supplement C, 20-23

Goh, P. P. (2008) ‘Status of Diabetic Retinopathy Among Diabetics Registered to the Diabetic Eye

Registry, National Eye Database, 2007’. Medical Journal Malaysia (63), 24-28

Gonzalez, R. C., and Woods, R. E. (2002) Digital Image Processing. 2nd edn. United States of America:

Prentice Hall

Gonzalez, R. C., Woods, R. E., and Eddins, S. L. (2009) Digital Image Processing Using MATLAB.

2nd edn. United States of America: Gatesmark Publishing

Harangi, B., Antal, B., and Hajdy, A. (2012) ‘Automatic Exudate Detection with Improved Naïve-Bayes

Classifier’. in Soda, P. (ed.) Proceedings of the 25th International Symposium on Computer-

Based Medical Systems (CBMS). held 20-22 June 2012 at Rome. USA: IEEE, 1-4

Hassan, S. S. A., and Bong, D. B. L. (2012) ‘Detection of Neovascularization in Diabetic Retinopathy’.

Journal Digital Imaging 25, 437-444

162

Hatanaka, Y., Inoue, T., Okumura, S., Muramatsu, C., and Fujita, H. (2012) ‘Automated Microaneurysm

Detection Method Based on Double-ring Filter and Feature Analysis in Retinal Fundus Images’.

in Soda, P. (ed.) Proceedings of the 25th International Symposium on Computer-Based Medical

Systems, CBMS. held 20-22 June 2012 at Rome. USA: IEEE, 1-4

Health Technology Assessment Unit, Medical Development Division, Ministry of Health Malaysia

(2002) Report Screening for Diabetic Retinopathy. Kuala Lumpur: Ministry of Health Malaysia

Hipwell, J. H., Strachant, F., Olson, J. A., McHardy, K. C., Sharp, P. F., and Forrester, J. V. (2000)

‘Automated Detection of Microaneurysms in Digital Red-free Photographs: A Diabetic

Retinopathy Screening Tool’. Diabetic Medicine 17, 588-594

Hoover, A., and Goldbaum, M. (2003) ‘Locating the Optic Nerve in a Retinal Image Using the Fuzzy

Convergence of the Blood Vessels’. IEEE Transactions on Medical Imaging 22 (8), 951-958

Hoover, A., Kouznetsova, V., and Goldbaum, M. (2000) ‘Locating Blood Vessels in Retinal Images by

Piece-wise Threshold Probing of a Matched Filter Response’. IEEE Transactions on Medical

Imaging 19 (3), 203-210

Hunter, A., Lowell, J. A., Steel, D., Ryder, B., Basu, A. (2011) ‘Automated Diagnosis of Referable

Maculopathy in Diabetic Retinopathy Screening’. in Proceedings of 2011 Annual International

Conference of the IEEE Engineering in Medicine and Biology Society, EMBC 2011. held 30

August- 3 September 2011 at Boston, Massachusetts. USA: IEEE, 3375-3378

Hutchins, E, Coppell, K. J., Morris, A., and Sanderson, G. (2012) ‘Diabetic Retinopathy Screening in

New Zealand Requires Improvement: Results from a Multi-centre Audit’. Australian and New

Zealand Journal of Public Health 36(3), 257-262

Ibrahim, H., and Kong, N. S. P. (2007) ‘Brightness Preserving Dynamic Histogram Equalization for

Image Contrast Enhancement’. IEEE Transactions on Consumer Electronics 53 (4), 1752-1758

ImageJ (2014) ImageJ An Open Platform for Scientific Image Analysis [online] available from

<http://imagej.net/ImageJ> [18 July 2014]

Imandoust, S. B., and Bolandraftar, M. (2013) ‘Application of k-nearest Neighbor (KNN) Approach for

Predicting Economic Events: Theoretical Background’. International Journal of Engineering

Research and Application 3 (5), 605-610

Itseez (2014) OpenCV [online] available from <http://opencv.org> [1 December 2013]

Jaafar, H. F., Nandi, A. K., and Al-Nuaimy, W. (2010) ‘Automated Detection of Exudates in Retinal

Images Using a Split-and-merge Algorithm’. in Proceedings of 18th European Signal

Processing Conference, EUSIPCO-2010. held 23-27 August 2010 at Aalborg. USA: IEEE,

1622-1626

Jakkula, V. (2008) Tutorial on Support Vector Machine (SVM). Pullman: Washington State University

Jimenez, S., Alemany, P., Benjumea, F. N., Serrano, C., Acha, B., Fondon, I., Carral, F., and Sanchez, C.

(2011) ‘Automatic Detection of Microaneurysms in Colour Fundus Images’. Arch Soc Esp

Oftalmol 86 (9), 277-281

Joshi. S., and Karule, P. T. (2012) ‘Retinal Blood Vessel Segmentation’. International Journal of

Engineering and Innovative Technology 1 (3), 175-178

163

Karasulu, B. (2012) ‘Automated Extraction of Retinal Blood Vessels: A Software Implementation’.

European Scientific Journal 8 (30), 47-57

Kauppi, T., Kalesnykiene, V., Kamarainen, J.-K., Lensu, L., Sorri, I., Uusitalo, H., Kalviainen, H., and

Pietila, J. (2006) DIARETDB0: Evaluation Database and Methodology for Diabetic Retinopathy

Algorithms. Finland: Lappeenranta University of Technology

Kauppi, T., Kalesnykiene, V., Kamarainen, J.-K., Lensu, L., Sorri, I., Raninen A., Voutilainen R.,

Uusitalo, H., Kalviainen H. and Pietila, J. (2007) DIARETDB1 Diabetic Retinopathy Database

and Evaluation Protocol. Finland: Lappeenranta University of Technology

Kose, C., Sevik, U., Ikibas, C., and Erdol, H. (2012) ‘Simple Methods for Segmentation and

Measurement of Diabetic Retinopathy Lesions in Retinal Fundus Images’. Computer Methods

and Programs in Biomedicine 107, 274-293

Kothari, C. R. (2004) Research Methodology Methods and Techniques. New Delhi: NewAge

International Publishers

Kumar, T. A., Priya, S., and Paul, V. (2013) ‘A Novel Approach to the Detection of Macula in Human

Retinal Imagery’. International Journal of Signal Processing Systems 1 (1), 23-28

Kwan, H. K. (2003) ‘Fuzzy Filters for Noisy Image Filtering’. in Proceedings of the 2003 International

Symposium on Circuits and Systems, 2003, ISCAS ’03. held 25-28 May 2003 at Bangkok,

Thailand. USA: IEEE, IV-161- IV-164

Lazar, I., and Hajdu, A. (2010) ‘Retinal Microaneurysm Detection Based on Intensity Profile Analysis’

in Proceedings of the 8th International Conference on Applied Informatics. held 27-30 January

2010 at Eger, Hungary. 157-165

Letchuman, G. R., Wan Nazaimoon, W. M., Wan Mohamad, W. B., Chandran, L. R., Tee, G. H.,

Jamaiyah, H., Isa, M. R., Zanariah, H., Fatanah, I., and Ahmad Faudzi, Y. (2010) ‘Prevalance of

Diabetes in the Malaysian National Health Morbidity Survey III 2006’. Medical Journal

Malaysia 65 (3), 173-179

Li, B. M., Chui, C. K., Chang, S., Ong, S. H. (2011) ‘Integrating Spatial Fuzzy Clustering with Level Set

Methods for Automated Medical Image Segmentation’. Computers in Biology and Medicine 41

(1), 1-10

Lichode, R. V., and Kulkarni, P. S. (2013) ‘Automatic Diagnosis of Diabetic Retinopathy by Hybrid

Multilayer Feed Forward Neural Network’. International Journal of Science, Engineering and

Technology Research (IJSETR) 2 (9), 1727-1733

Lim, G, Lee, M. L., Hsu, W., and Wong, T. Y. (2014) ‘Transformed Representations for Convolutional

Neural Networks in Diabetic Retinopathy Screening. Modern Artificial Intelligent for Health

Analytics, 21-25

Lin, A., Hoffman, D., Gaasterland, D, E., and Caprioli, J. (2003) ‘Neural Networks to Identify

Glaucomatous Visual Field Progression’. American Journal of Ophthalmology 135 (1), 49-54

Mafauzy, M., Hussein, Z., and Chan, S. P. (2011) ‘The Status of Diabetes Control in Malaysia: Results of

DiabCare 2008’. Medical Journal Malaysia 66 (3), 175-181

164

Malhotra, N. K., and Birks D. F. (2006) Marketing Research An Applied Approach. 2nd edn. England:

Pearson Education Limited

Mallika, P. S., Lee, P. Y., Cheah, W. L., Wong, J. S., Syed Alwi, S. A. R., Nor Hayati, H., and Tan, A. K.

(2011) ‘Risk Factors for Diabetic Retinopathy in Diabetics Screened Using Fundus Photography

at a Primary Health Care Setting in East Malaysia’. Malaysian Family Physician 2011 6 (2-3),

60-65

Martins, C. I. O., Veras, R. M, S., Ramalho, G. L. B., Medeiros, F. N. S., and Ushizima, D. M. (2010).

‘Automatic Microaneurysms Detection and Characterization Through Digital Color Fundus

Images’. in Proceedings of the International Joint Conference – Brazilian Symposium on

Artificial Intelligence and Brazilian Symposium on Neural Networks – II Workshop on

Computational Intelligence. held 23-28 October 2010 at Salvador, Bahia-Brazil. USA: Lawrence

Berkeley National Laboratory, 1-6

Mathers, C. D., and Loncar, D. (2006) ‘Projections of global mortality and burden of disease from 2002

to 2030’. PLoS Med 3(11), e442

MathWorks (2014) MATLAB [online] available from <http://www mathworks.co.uk/products/matlab/>

[1 November 2013]

Messidor (2004) Messidor: Digital Retinal Images [online] available from <http://messidor.crihan.fr/

index-en.php> [16 January 2014]

Ministry of Health Diabetic Retinopathy Screening Team 2012 (2012a) Diabetes Mellitus and

Complications –Module 2-2012. Putrajaya: Ministry of Health Malaysia

Ministry of Health Diabetic Retinopathy Screening Team 2012 (2012b) Handbook Guide to Diabetic

Retinopathy Screening -Module 5-2012. Putrajaya: Ministry of Health Malaysia

Ministry of Health Diabetic Retinopathy Screening Team 2012 (2012c) Diabetes Mellitus and Diabetic

Retinopathy-An Overview -Module 1-2012. Putrajaya: Ministry of Health Malaysia

Ministry of Health Diabetic Retinopathy Screening Team 2012 (2012d) Introduction to Fundus Camera–

Module 4-2012. Putrajaya: Ministry of Health Malaysia

Ministry of Health Malaysia, Malaysian Society of Ophthalmology and Academy of Medicine of

Malaysia (2011) Clinical Practice Guidelines Screening of Diabetic Retinopathy. Putrajaya:

Ministry of Health Malaysia

Mizutani, A., Muramatsu, C., Hatanaka, Y. (2009) ‘Automated Microaneurysm Detection Method Based

on Double-ring Filter in Retinal Fundus Images’. Medical Imaging 2009 Computer-Aided

Diagnosis, 1-8

Mookiah, M. R. K., Acharya, U. R., Martis, R. J., Chua, C.K., Lim, C. M., Ng, E. Y. K., and Laude, A.

(2013) ‘Evolutionary Algorithm Based Classifier Parameter Tuning for Automatic Diabetic

Retinopathy Grading: A Hybrid Feature Extraction Approach’. Knowledge-Based Systems 39, 9-

22

Mookiah, M. R. K., Rajendra, U. A., Lim, C. M., Petznick, A., and Suri, J. S. (2012) ‘Data Mining

Technique for Automated Diagnosis of Glaucoma Using Higher Order Spectra and Wavelet

Energy Features’. Knowledge-Based Systems 33, 73-82

165

Mubbashar, M., Usman, A., and Akram, M. U. (2011) ‘Automated System for Macula Detection in

Digital Retinal Images’. in Proceedings of the 2011 International Conference on Information

and Communication Technologies, ICICT. held 23-24 July 2011at Karachi. USA: IEEE, 1-5

Murugan, R., Korah, R., Nasreen, F. S., and Ventaka, H. T. (2003) ‘Microaneurysms Detection Methods

In Retinal Images Using Mathematical Morphology’. International Journal of Advances in

Engineering Science and Technology 2 (1), 120-128

Nagaveena, Deepashree, D., and Kumar S. C. P. (2013) ‘Vessels Segmentation in Diabetic Retinopathy

by Adaptive Median Thresholding’. The International Journal of Science & Technoledge 1 (1),

17-22

National Institute for Clinical Excellence (2002) Management of Type 2 Diabetes. Retinopathy-screening

and Early Nanagement. London: NICE

NED Steering Committee Members (2015) The 7th Report of the National Eye Database 2013 [online]

available from <http://www.acrm.org. my/ned/NEDreport/NED_2013.pdf> [14 December 2015]

NHS Choices (2012) Diabetes [online] available from <http://conditions/diabetes/ pages/ diabetes.aspx>

[1 July 2013]

Niemeijer, M., Ginneken, –van B., Staal, J., Suttorp-Schulten, M. S. A., and Abramoff, M. D. (2005)

‘Automatic Detection of Red Lesions in Digital Color Fundus Photographs’. IEEE Transactions

on Medical Imaging 24 (5), 584-592

Niemeijer, M., Ginnerken, -van B., Cree, M. J., Mizutani, A., Quellec, G., Sanchez, C. I., Zhang, B.,

Hornero, R., Lamard, M., Muramatsu, C., Wu, X., Cazuquel, G., You, J., Mayo, A., Li, Q.,

Hatanaka, Y., Cochener, B., Roux, C., Karray, F., Garcia, M., Fujita, H., Abramoff, M. D.

(2010) ‘Retinopathy Online Challenge: Automatic Detection of Microaneurysms in Digital

Color Fundus Photographs’. IEEE Transactions on Medical Imaging 29 (1), 185-195

Noronha, K., Nayak, J., and Bhat, S. N. (2006) ‘Enhancement of Retinal Fundus Image to Highlight the

Features for Detection of Abnormal Eyes’. in Proceedings of the 2006 IEEE Region 10

Conference ,TENCON2006. held 14-17 November 2006 at Hong Kong. USA:IEEE, 1-4

Osareh, A., Mirmehdi, M., Thomas, B., and Markham, R. (2003) ‘Automated Identification of Diabetic

Retinal Exudates in Digital Colour Images’. British Journal of Ophthalmology 87, 1220-1223

Patil, J., and Chaudhari, A. L. (2012) ‘Development of Digital Image Processing Using Fuzzy Gaussian

Filter Tool for Diagnosis of Eye Infection’. International Journal of Computer Applications

51 (19), 10-12

Prakash. J., and Sumanthi, K. (2013) ‘Detection and Classification of Microaneurysms for Diabetic

Retinopathy’. International Journal of Engineering Research and Applications, 31-36

Priya, R., and Aruna, P. (2011) ‘Review of Automated Diagnosis of Diabetic Retinopathy Using the

Support Vector Machine’. International Journal of Applied Engineering Research 1 (4), 844-863

Priya, R and Aruna, P. (2012) ‘SVM and Neural Network Based Diagnosis of Diabetic Retinopathy’.

International Journal of Computer Applications 41 (1), 6-12

Priya, R and Aruna, P. (2013a) ‘Diagnosis of Diabetic Retinopathy Using Machine Learning Techniques’.

Journal on Soft Computing 3 (4), 563-575

166

Priya, R., Aruna, P., and Suriya, R. (2013b) ‘Image Analysis Technique for Detecting Diabetic

Retinopathy’. International Journal of Computer Applications 1, 34-38

Punnolil, A. (2013) ‘A Novel Approach for Diagnosis and Severity Grading of Diabetic Maculopathy. in

Proceedings of the 2013 International Conference on Advances in Computing, Communications

and Informatics. held 22-25 August 2013 at Mysore, India. New York: IEEE, 1230-1235

Quellec, G., Lamard, M., Josselin, P. M., Cazuguel, G., Cochener, B., Roux, C. (2008) ‘Optimal Wavelet

Transform for the Detection of Microaneurysms in Retina Photographs’. IEEE Transactions on

Medical Imaging 27 (9), 1230-1241

Rahim, S, S., Jayne, C., Palade, V., and Shuttleworth, J. (2015a) ‘Automatic Detection of

Microaneurysms in Colour Fundus Images for Diabetic Retinopathy Screening’. Journal of

Neural Computing and Applications 521, 1-16

Rahim, S. S., Palade, V., Jayne, C., Holzinger, A., and Shuttleworth, J. (2015c) ‘Detection of Diabetic

Retinopathy and Maculopathy in Eye Fundus Images Using Fuzzy Image Processing’. in Guo,

Y. et al. (ed.) Proceedings of 8th

International Conference on Brain Informatics and Health, BIH

2015, LNAI 9250. held 30 August-2 September 2015 at London, UK. Switzerland: Springer,

379-388

Rahim, S. S., Palade, V., Shuttleworth, J., and Jayne, C. (2014) ‘Automatic Screening and Classification

of Diabetic Retinopathy Fundus Images’. in Mladenov, V. et al. (ed.) Proceedings of 15th

International Conference on Engineering Applications of Neural Networks, EANN 2014,

Communications in Computer and Information Science 459. held 5-7 September 2014 at Sofia,

Bulgaria. Switzerland: Springer, 113-122

Rahim, S. S., Palade, V., Shuttleworth, J., and Jayne, C. (2016) ‘Automatic Screening and Classification

of Diabetic Retinopathy and Maculopathy Using Fuzzy Image Processing’. in Zhong, N., and

Peng, H. (ed.) Brain Informatics 40708, 1-19

Rahim, S. S., Palade, V., Shuttleworth, J., Jayne, C., and Raja Omar, R. N. (2015b) ‘Automatic Detection

of Microaneurysms for Diabetic Retinopathy Screening Using Fuzzy Image Processing’. in

Iliadis, L., and Jayne, C. (ed.) Proceedings of 16th

International Conference on Engineering

Applications of Neural Networks, EANN 2015, Communications in Computer and Information

Science 517. held 25-28 September 2015 at Rhodes, Greece. Switzerland: Springer, 69-79

Ravishankar, A., Jain, A., and Mittal, A. (2009) ‘Automated Feature Extraction or Early Detection of

Diabetic Retinopathy in Fundus Images’. in Proceedings of the IEEE Conference on Computer

Vision and Pattern Recognition 2009, CVPR 2009. held 20-25 June 2009 at Miami, FL. USA:

IEEE, 210-217

Rhody, H. (2005) Lecture 10: Hough Circle Transform,[online lecture] DIP Lecture 10, 11 October 2005.

Rochester: Rochester Institute of Technology. available from <http://www.cis rit.edu/

class/simg782/ lectures/lecture_10/lec782_05_10.pdf> [30 June 2014]

Rokash, L., and Maimon, O. (2005) Decision Trees. The Data Mining and Knowledge Discovery

Handbook. 1st edn. Springer Science and Business Media: New York

167

Saleh, M. D., and Eswaran, C. (2012) ‘An Automated Decision-support System for Non-Proliferative

Diabetic Retinopathy Disease Based on Mas and HAs Detection’. Computer Methods and

Programs in Biomedicine 108, 186-196

Sanchez, C. I., Hornero, R., Mayo, A., and Garcia, M. (2009) ‘Mixture Model-based Clustering and

Logistic Regression for Automatic Detection of Microaneurysms in Retinal Images’. in

Karssemeijer, N., and Giger, M. L. (ed.) Proceedings of the SPIE 7260, Medical Imaging 2009-

Computer-Aided Diagnosis. held 7 February 2009 at Lake Buena Vista, FL. USA: SPIE

Scanlon, P. H., Wilkinson, C. P., Aldington, S. J., and Matthews, D. R. (2009) A Practical Manual of

Diabetic Retinopathy Management. Chicester, UK: Wiley-Blackwell

Sekhar, S., Al-Nuaimy, W., and Nandi, A. K. (2008) ‘Automated Localisation of Optic Disk and Fovea in

Retinal Fundus Images’. in Proceedings of the 16th European Signal Processing Conference.

held 25-29 August 2008 at Lausanne, Switzerland. USA: IEEE, 1-5

Selvathi, D., Prakash, N. B., and Balagopal, N. (2012) ‘Automated Detection of Diabetic Retinopathy for

Early Diagnosis Using Feature Extraction and Support Vector Machine’. International Journal

of Emerging Technology and Advanced Engineering 2 (11), 762-767

Sheet, D., Garud, H., Suveer, A., Mahadevappa, M., and Chatterjee, J. (2010) ‘Brightness Preserving

Dynamic Fuzzy Histogram Equalization’. IEEE Transactions on Consumer Electronics 56 (4),

2475-2480

Shome, S. K., and Vadali, S. R. K. (2011) ‘Enhancement of Diabetic Retinopathy Imagery Using Contrast

Limited Adaptive Histogram Equalization’. International Journal of Computer Science and

Information Technologies 2 (6), 2694-2699

Siddalingaswamy, P. C., and Prabhu, K. G. (2010) ‘Automatic Grading of Diabetic Maculopathy Severity

Levels’. in Mahadevappa, M. et al. (ed.) Proceedings of the 2010 International Conference on

Systems in Medicine and Biology. held 16-18 December 2010 at Kharagpur, India. New Delhi:

Excel India Publishers, 331-334

Sivaprasad, S., Gupta, B., Crossby-Nwaobi, R., and Evans, J. (2012) ‘Prevalence of Diabetic Retinopathy

In Various Ethnic Groups: A Worldwide Perspective’. Survey of Ophthalmology 57 (4), 347-370

Social Dimensions of Watershed Planning (2006) Conducting a Social Profile [online] available from

<http://www.watershedplanning.illinois.edu/profile_steps/step3. cfm> [16 July 2013]

Solberg, A. (2009) Hough Transform [online lecture] INF 4300, 21 October 2009. Norway: University of

Oslo. available from <http://www.uio no/studier/emner/matnat/ifi/INF4300/h09/

undervisningsmateriale/hough09.pdf> [11 August 2014]

Sonka, M., and Fitzpatrick, J. M. (2000) Handbook of Medical Imaging Volume 2. Medical Image

Processing and Analysis. United States of America: SPIE Press

Sonka, M., Hlavac, V., and Boyle, R. (2008) Image Processing, Analysis, and Machine Vision. United

States of America: Cengage Learning

168

Sopharak, A., New, K.T., Moe, Y. A., Dailey, M. N., and Uyyanonvara, B. (2008) ‘Automatic Exudate

Detection with a Naïve Bayes Classifier’. in Proceedings of the 2008 International Conference

on Embedded Systems and Intelligent Technology. held 27-29 February 2008 at Bangkok,

Thailand. Thailand: IEEE, 139-142

Sopharak, A., Uyyanonvara, B., and Barman, S. (2009a) ‘Automatic Exudate Detection for Diabetic

Retinopathy Screening’. ScienceAsia 35, 80-88

Sopharak, A., Uyyanonvara, B., and Barman, S. (2009b) ‘Automatic Exudate Detection from Non-dilated

Diabetic Retinopathy Retinal Images Using Fuzzy C-means Clustering’. Sensors 2009, 2148-

2161

Sopharak, A., Uyyanonvara, B., and Barman, S. (2013) ‘Automated Microaneurysm Detection

Algorithms Applied to Diabetic Retinopathy Retinal Images’. Maejo International Journal of

Science and Technology 7 (2), 294-314

Sopharak, A., and Uyyanonvara, B. (2006) ‘Automatic Exudates Detection on Thai Diabetic Retinopathy

Patients’ Retinal Images’. in Proceedings of the 2006 International Conference on Electrical

Engineering/Electronics, Computer, Telecommunications and Information Technology (ECTI).

held 10-13 May 2006 at Ubon Ratchathani, Thailand. Thailand: IEEE, 709-712

Sopharak, A., and Uyyanonvara, B. (2007) ‘Automatic Exudates Detection from Non-dilated Diabetic

Retinopathy Retinal Image Using Fuzzy C-means Clustering’. in Proceedings of the 3rd WACBE

World Congress on Bioengineering 2007,WACBE 2007. held 9-11 July 2007 at Bangkok,

Thailand. 1-4

Sopharak, A., Uyyanonvara, B., Barman, S., and Williamson, T. H. (2007) ‘Automatic Exudates

Detection from Diabetic Retinopathy Retinal Image Using Fuzzy C-means and Morphological

Methods’. in Proceedings of the Third IASTED International Conference Advances in Computer

Science and Technology. held 2-4 April 2007 at Phuket, Thailand. Anaheim, CA, USA: ACTA

Press, 359-364

Sopharak, A., Uyyanonvara, B., Barman, S., and Williamson, T. H. (2008) ‘Automated Detection of

Diabetic Retinopathy Exudates from Non-dilated Retinal Images Using Mathematical

Morphology’. Computerized Medical Imaging and Graphics 32, 720-727

Sopharak, A., Uyyanonvara, B., Barman, S., and Williamson, T. H. (2010) ‘Machine Learning Approach

To Automatic Exudate Detection in Retinal Images from Diabetic Patients’. Journal of Modern

Optics 5 (2), 124-135

Sounders, M., Lewis, P., and Thornhill, A. (2009) Research Methods for Business Students. 5th edn.

England: Pearson Education Limited

Staal, J. J., Abramoff, M. D., Niemeijer, M., Viergever, M. A., and Ginneken- van, B. (2004) ‘Ridge

Based Vessel Segmentation in Color Images of the Retina’. IEEE Transactions on Medical

Imaging 23, 501-509

Streeter, L., and Cree, M. J. (2003) ‘Microaneurysm Detection in Colour Fundus Images’. Image and

Vision Computing NZ, 280-285

169

Sujithkumar, S. B., and Vipula, S. (2012) ‘Automatic Detection of Diabetic Retinopathy in Non-dilated

RGB Retinal Fundus Images’. International Journal of Computer Applications 47 (19), 26-32

Sundhar, C., and Archana, D. (2014) ‘Automatic Screening of Fundus Images for Detection of Diabetic

Retinopathy’. International Journal of Communication and Computer Technologies 2 (1), 100-

105

Tajunisah, I., Nabilah, H., and Reddy, S. C. (2006) ‘Prevalence and risk factors for diabetic retinopathy –

A study of 217 Patients from University of Malaya Medical Centre’. Medical Journal Malaysia

61 (4), 451-456

Tariq, A., Akram, M. U., Shaukat, A., Khan, S. A. (2013) ‘Automated Detection and Grading of Diabetic

Maculopathy in Digital Retinal Images’. Journal of Digital Imaging 2013 26 (4), 803-812

Taylor, R., and Batey, D. (2012) Handbook of retinal screening in diabetes: diagnosis and management.

Chichester England: John Wiley & Sons, Ltd

The MathWorks, Inc. (2016a) Support Vector Machines (SVM) [online] available from

<http://uk.mathworks.com/help/stats/support-vector-machines-svm.html> [1 August 2014]

The MathWorks, Inc. (2016b) Naïve Bayes Classification [online] available from

<http://uk.mathworks.com/help/stats/naive-bayesclassification.html?searchHighlight=

naive%20bayes> [1 August 2014]

The MathWorks, Inc. (2016c) Fuzzy Logic Image Processing [online] available from

<http://uk.mathworks.com/help/fuzzy/examples/fuzzy-logic-image-processing html>

[23 February 2015]

Thevi, T., Basri, M., and Reddy, S. C. (2012) ‘Prevalence of Eye Diseases and Visual Impairment Among

the Rural Population- A Case Study of Temerloh Hospital’. Malaysian Family Physician 7 (1),

6-10

Toh, K. K. V., and Mat Isa, N. A. (2010) ‘Noise Adaptive Fuzzy Switching Median Filter for Salt-and-

pepper Noise Reduction’. IEEE Signal Processing Letters 17 (3), 281-284

Verma, K., Deep, P., and Ramakrishnan, A. G. (2011) ‘Detection and Classification of Diabetic

Retinopathy Using Retinal Images’. in Negi, A. et al (ed.) Proceedings of the 2011 Annual IEEE

India Conference ,INDICON-2011, ‘Engineering Sustainable Solutions’. held 16-18 December

2011 at Hyderabad, India. Hyderabad: IEEE, 1-6

Vidyasari, R., Sovani, I., Mengko, T. L. R., and Zakaria, H. (2011) ‘Vessel Enhancement Algorithm in

Digital Retinal Fundus Microaneurysms Filter for Nonproliferative Diabetic Retinopathy

Classification’. in Proceedings of the 2011 2nd International Conference in Instrumentation,

Communication, Information Technology and Biomedical Engineering, ICICI-BME. held 8-9

November 2011 at Bandung, Indonesia. USA: IEEE, 278-281

Vimala, A. G. S. G., and Kajamohideen, S. (2014) ‘Detection of Diabetic Maculopathy in Human Retinal

Images Using Morphological Operations’. Online Journal of Biological Sciences 14,175-180

170

Walter, T., and Klein, J. -C. (2002) ‘Automatic Detection of Microaneurysms in Color Fundus Images of

The Human Retina by Means of the Bounding Box Closing’. in Colosimo, A. et. al. (ed.)

Proceedings of the Third International Symposium on Medical Data Analysis, ISMDA 2002,

LNCS 2526. held 8-11 October 2002 at Rome, Italy. Berlin, Heidelberg: Springer-Verlag, 210-

220

Walter, T., Massin, P., Erginay, A., Ordonez, R., Jeulin, C., and Klein, J. -C. (2007) ‘Automatic Detection

of Microaneurysms in Color Fundus Images’. Medical Image Analysis 11 (6), 555-566

Wilkinson, C. P., Ferris, F. L., Klein, R. E., Lee, P. P., Agardh, C. D., Davis, M., Dills, D., Kampik, A.,

Pararajasegaram, R., and Verdaguer, J. T. (2003) ‘Proposed International Clinical Diabetic

Retinopathy and Diabetic Macula Edema Disease Severity Scales’. American Academy of

Ophthalmology 110 (9), 1677-1682

Wisaeng, K., Hiransakolwong, N., and Pothiruk, E. (2012) ‘Automatic Detection of Exudates in Diabetic

Retinopathy Images. Journal of Computer Science 8 (8), 1304-1313

Wisaeng, K., Hiransakolwong, N., and Pothiruk, E. (2013) ‘Automatic Detection of Exudates in Digital

Retinal Images’. International Journal of Computer Applications 64 (4), 19-26

Wisaeng, K., Hiransakolwong, N., and Pothiruk, E. (2014) ‘Automatic Detection of Optic Disc in Digital

Retinal Images’. International Journal of Computer Applications 90 (50), 15-20

World Health Organization (2005) Prevention of Blindness from Diabetes Mellitus, Report of WHO

Consultation. Geneva: WHO

World Health Organization (2012a) Global status report on noncommunicable diseases 2014. Geneva:

WHO

World Health Organization (2012b) Global data on visual impairments 2010. Geneva:WHO

Yadao, P., Naval, S., and Maheshwari, A. (2014) ‘Blood Vessel Segmentation in Retinal Fundus Images

Using Matched Filter’. Journal of Harmonized Research in Engineering 2(1), 58-61

Zadeh, L. A. (1965) ‘Fuzzy Sets’. Information and Control 8, 338-353

Zhang, B., Wu, X., You, J., Li, Q., and Karray, F. (2009) ‘Hierarchical Detection of Red Lesions in

Retinal Images by Multiscale Correlation Filtering’. in Karssemeijer, N., and Giger, M. L. (ed.)

Proceedings of the SPIE 7260, Medical Imaging 2009-Computer-Aided Diagnosis. held 7

February 2009 at Lake Buena Vista, FL. USA: SPIE

Zhang, B., Zhang, L., Zhang, L., and Karray, F. (2010) ‘Retinal Vessel Extraction by Matched Filter with

First-Order Derivative of Gaussian’. Computers in Biology and Medicine 40, 438-445

Zhang, B., Wu, X., You, J., Li, Q., and Karray, F. (2010) ‘Detection of Microaneurysms Using Multi-

Scale Correction Coefficients’. Pattern Recognition 43, 2237-2248

Zhang, X., and Chutatape, O. (2005) ‘A SVM Approach for Detection of Hemorrhages in Background

Diabetic Retinopathy’. in Proceedings of International Joint Conference on Neural Networks.

held 31 July- 4 August 2005 at Montreal, Canada. USA: IEEE, 2435-2440

Zohra, B. F., and Mohamed, B. (2009) ‘Automated Diagnosis of Retinal Images Using the Support

Vector Machine (SVM)’. Faculte des Science. Department of Informatique, USTO, Algerie

171

APPENDICES

Appendix A Data Collection

A.1 Data Collection Process

Figure A.1 Data collection and diabetic retinopathy screening process

Eye Clinic, Melaka Hospital, Malaysia

Fundus Camera VX-10

Screening Process

This item has been removed due to Data Protection. The unabridged version of the thesis can be found in the


172

No Diabetic Retinopathy

Figure A.2 No diabetic retinopathy images

Mild DR without maculopathy

Figure A.3 Mild DR without maculopathy images

173

Mild DR with maculopathy

Figure A.4 Mild DR with maculopathy images

Moderate DR without maculopathy

Figure A.5 Moderate DR without maculopathy images

174

Moderate DR with maculopathy

Figure A.6 Moderate DR with maculopathy images

Severe DR without maculopathy

Figure A.7 Severe DR without maculopathy images

175

Severe DR with maculopathy

Figure A.8 Severe DR with maculopathy images

Proliferative DR without maculopathy

Figure A.9 Proliferative DR without maculopathy images

176

Proliferative DR with maculopathy

Figure A.10 Proliferative DR with maculopathy images

Advanced Diabetic Eye Disease (ADED)

Figure A.11 Advanced diabetic eye disease images

184

Appendix B Online Developed Dataset Screenshots

B.1 Website Screenshots (http://creative.coventry.ac.uk/fundus)

Figure B.1 ‘Home’ page

Figure B.2 ‘About’ page





185

Figure B.3 ‘Team’ page

Figure B.4 ‘Publications’ page





187

Appendix C Research Ethics Approval

Automatic Screening and Classification of Diabetic ...

Documents

Transcript of Automatic Screening and Classification of Diabetic ...